[pre-commit.ci] pre-commit suggestions

.pre-commit-config.yaml

@@ -38,7 +38,7 @@ repos:
       - id: detect-private-key
 
   - repo: https://github.com/codespell-project/codespell
-    rev: v2.4.2
+    rev: v2.4.3
     hooks:
       - id: codespell
         additional_dependencies: [tomli]
@@ -71,19 +71,19 @@ repos:
         args: ["--print-width=140"]
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.15.9
+    rev: v0.16.10
     hooks:
       - id: ruff
         args: ["--fix"]
       - id: ruff-format
       - id: ruff
 
   - repo: https://github.com/tox-dev/pyproject-fmt
-    rev: v2.21.0
+    rev: v2.30.1
     hooks:
       - id: pyproject-fmt
         additional_dependencies: [tox]
   - repo: https://github.com/abravalheri/validate-pyproject
-    rev: v0.25
+    rev: "0.26"
     hooks:
       - id: validate-pyproject

README.md

@@ -282,9 +282,9 @@ Test the server in a separate terminal and integrate the model API into your AI
 ```python
 # 3) Use the server (in a separate Python session)
 import requests, json
+
 response = requests.post(
-    "http://127.0.0.1:8000/predict",
-    json={"prompt": "Fix typos in the following sentence: Example input"}
+    "http://127.0.0.1:8000/predict", json={"prompt": "Fix typos in the following sentence: Example input"}
 )
 print(response.json()["output"])
 ```

pyproject.toml

@@ -22,6 +22,7 @@ classifiers = [
   "Programming Language :: Python :: 3.12",
   "Programming Language :: Python :: 3.13",
   "Programming Language :: Python :: 3.14",
+  "Programming Language :: Python :: 3.15",
 ]
 dependencies = [
   # download models:
@@ -89,15 +90,15 @@ urls.homepage = "https://github.com/lightning-AI/litgpt"
 scripts.litgpt = "litgpt.__main__:main"
 
 [tool.setuptools]
-package-data.litgpt = [
-  "LICENSE.md",
-  "README.md",
-]
 packages.find.include = [
   "litgpt",
   "litgpt.*",
 ]
 packages.find.exclude = []
+package-data.litgpt = [
+  "LICENSE.md",
+  "README.md",
+]
 
 [tool.ruff]
 target-version = "py310"
@@ -114,12 +115,6 @@ lint.select = [
   "UP", # see: https://docs.astral.sh/ruff/rules/#pyupgrade-up
   "W",  # see: https://pypi.org/project/pycodestyle
 ]
-# extend-select = [
-#    "C4",  # see: https://pypi.org/project/flake8-comprehensions
-#    "PT",  # see: https://pypi.org/project/flake8-pytest-style
-#    "RET",  # see: https://pypi.org/project/flake8-return
-#    "SIM",  # see: https://pypi.org/project/flake8-simplify
-# ]
 lint.ignore = [
   "E501", # Line too long
   "E731", # Do not assign a lambda expression, use a def
@@ -128,14 +123,20 @@ lint.ignore = [
 ]
 # Use Google-style docstrings.
 lint.pydocstyle.convention = "google"
+# extend-select = [
+#   "C4",  # see: https://pypi.org/project/flake8-comprehensions
+#   "PT",  # see: https://pypi.org/project/flake8-pytest-style
+#   "RET", # see: https://pypi.org/project/flake8-return
+#   "SIM", # see: https://pypi.org/project/flake8-simplify
+# ]
 
 [tool.codespell]
-# skip = '*.py'
-quiet-level = 3
 ignore-words-list = """
   tral, \
   Rockerfeller
 """
+# skip = "*.py"
+quiet-level = 3
 
 [tool.pytest]
 ini_options.addopts = [

tutorials/convert_lit_models.md

@@ -27,9 +27,7 @@ from transformers import AutoModel
 
 
 state_dict = torch.load("output_dir/model.pth")
-model = AutoModel.from_pretrained(
-    "output_dir/", local_files_only=True, state_dict=state_dict
-)
+model = AutoModel.from_pretrained("output_dir/", local_files_only=True, state_dict=state_dict)
 ```
 
 Alternatively, you can also load the model without copying the `config.json` file as follows:
@@ -107,7 +105,7 @@ litgpt convert_from_litgpt $finetuned_dir/final/ out/hf-tinyllama/converted
 import torch
 from transformers import AutoModel
 
-state_dict = torch.load('out/hf-tinyllama/converted/model.pth')
+state_dict = torch.load("out/hf-tinyllama/converted/model.pth")
 model = AutoModel.from_pretrained("TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T", state_dict=state_dict)
 ```
 

tutorials/deploy.md

@@ -35,8 +35,7 @@ You can now send requests to the inference server you started in step 2. For exa
 import requests, json
 
 response = requests.post(
-    "http://127.0.0.1:8000/predict",
-    json={"prompt": "Fix typos in the following sentence: Example input"}
+    "http://127.0.0.1:8000/predict", json={"prompt": "Fix typos in the following sentence: Example input"}
 )
 
 print(response.json()["output"])
@@ -63,9 +62,7 @@ Then, use the following updated code to query the inference server:
 import requests, json
 
 response = requests.post(
-    "http://127.0.0.1:8000/predict",
-    json={"prompt": "Fix typos in the following sentence: Example input"},
-    stream=True
+    "http://127.0.0.1:8000/predict", json={"prompt": "Fix typos in the following sentence: Example input"}, stream=True
 )
 
 # stream the response
@@ -121,14 +118,11 @@ from openai import OpenAI
 # Configure the client to use your local LitGPT server
 client = OpenAI(
     base_url="http://127.0.0.1:8000/v1",
-    api_key="not-needed"  # LitGPT doesn't require authentication by default
+    api_key="not-needed",  # LitGPT doesn't require authentication by default
 )
 
 response = client.chat.completions.create(
-    model="SmolLM2-135M-Instruct",
-    messages=[
-        {"role": "user", "content": "Hello! How are you?"}
-    ]
+    model="SmolLM2-135M-Instruct", messages=[{"role": "user", "content": "Hello! How are you?"}]
 )
 
 print(response.choices[0].message.content)

tutorials/developer-docs/adding-models.md

@@ -41,23 +41,25 @@ For example, suppose an entry for Llama 3 8B already exists and you want to add
 Copy the Llama 3 8B entry:
 
 ```python
- # https://huggingface.co/meta-llama/Meta-Llama-3-8B/blob/main/config.json
- dict(
-     name="Llama-3-8B{}",
-     hf_config=dict(org="meta-llama", name="Meta-Llama-3-8B{}"),
-     vocab_size=128256,
-     padding_multiple=64,
-     n_layer=32,
-     n_head=32,
-     n_query_groups=8,
-     rotary_percentage=1.0,
-     parallel_residual=False,
-     bias=False,
-     norm_class_name="RMSNorm",
-     mlp_class_name="LLaMAMLP",
-     intermediate_size=14336,
-     rope_base=500000,
- ),
+# https://huggingface.co/meta-llama/Meta-Llama-3-8B/blob/main/config.json
+(
+    dict(
+        name="Llama-3-8B{}",
+        hf_config=dict(org="meta-llama", name="Meta-Llama-3-8B{}"),
+        vocab_size=128256,
+        padding_multiple=64,
+        n_layer=32,
+        n_head=32,
+        n_query_groups=8,
+        rotary_percentage=1.0,
+        parallel_residual=False,
+        bias=False,
+        norm_class_name="RMSNorm",
+        mlp_class_name="LLaMAMLP",
+        intermediate_size=14336,
+        rope_base=500000,
+    ),
+)
 ```
 
 Then create the entry for the 70B model. Here, make sure you update the values according to the `config.json` file available on the HF hub:
@@ -130,21 +132,21 @@ If you are adding a new model class, find out its prompt style. First, check [li
 
 ```python
 class Llama3(PromptStyle):
-     def apply(self, prompt: str, **kwargs: str) -> str:
-         # https://github.com/meta-llama/llama3/blob/359887376f0aaf30e433f23e25df858d8c2a9833/llama/tokenizer.py#L202-L229
-         return (
-             "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n"
-             "You are a helpful assistant.<|eot_id|>\n"  # The system prompt is optional
-             "<|start_header_id|>user<|end_header_id|>\n\n"
-             f"{prompt}<|eot_id|>\n"
-             "<|start_header_id|>assistant<|end_header_id|>\n\n"
-         )
-
-     def stop_tokens(self, tokenizer: "Tokenizer") -> Tuple[List[int], ...]:
-         return (
-             [tokenizer.eos_id],
-             [tokenizer.token_to_id("<|eot_id|>")],
-         )
+    def apply(self, prompt: str, **kwargs: str) -> str:
+        # https://github.com/meta-llama/llama3/blob/359887376f0aaf30e433f23e25df858d8c2a9833/llama/tokenizer.py#L202-L229
+        return (
+            "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n"
+            "You are a helpful assistant.<|eot_id|>\n"  # The system prompt is optional
+            "<|start_header_id|>user<|end_header_id|>\n\n"
+            f"{prompt}<|eot_id|>\n"
+            "<|start_header_id|>assistant<|end_header_id|>\n\n"
+        )
+
+    def stop_tokens(self, tokenizer: "Tokenizer") -> Tuple[List[int], ...]:
+        return (
+            [tokenizer.eos_id],
+            [tokenizer.token_to_id("<|eot_id|>")],
+        )
 ```
 
 If your model requires a different prompt template, create a new `PromptStyle` class.

tutorials/developer-docs/python-api.md

@@ -74,12 +74,7 @@ dataset = llm.prepare_dataset(
 
 
 ```python
-llm.instruction_finetune(
-    config=None,
-    dataset=dataset,
-    max_iter=10,
-    method="full | lora | adapter | adapter_v2"
-)
+llm.instruction_finetune(config=None, dataset=dataset, max_iter=10, method="full | lora | adapter | adapter_v2")
 ```
 
 ```python
@@ -100,8 +95,7 @@ Then in another Python session:
 import requests, json
 
 response = requests.post(
-    "http://127.0.0.1:8000/predict",
-    json={"prompt": "Fix typos in the following sentence: Example input"}
+    "http://127.0.0.1:8000/predict", json={"prompt": "Fix typos in the following sentence: Example input"}
 )
 
 print(response.json()["output"])

tutorials/evaluation.md

@@ -112,15 +112,11 @@ Suppose you have a test dataset with the following structure:
 
 ```python
 test_data = [
-    {
-        "instruction": "Name the author of 'Pride and Prejudice'.",
-        "input": "",
-        "output": "Jane Austen."
-    },
+    {"instruction": "Name the author of 'Pride and Prejudice'.", "input": "", "output": "Jane Austen."},
     {
         "instruction": "Pick out the adjective from the following list.",
         "input": "run, tall, quickly",
-        "output": "The correct adjective from the list is 'tall.'"
+        "output": "The correct adjective from the list is 'tall.'",
     },
 ]
 ```
@@ -173,7 +169,7 @@ Next, we use a second LLM to calculate the response quality on a scale from 0 to
 
 
 ```python
-del llm # delete previous `llm` to free up GPU memory
+del llm  # delete previous `llm` to free up GPU memory
 scorer = LLM.load("meta-llama/Meta-Llama-3-8B-Instruct", access_token="...")
 ```
 
@@ -207,7 +203,7 @@ def generate_model_scores(data_dict, model, response_field="response", target_fi
 scores = generate_model_scores(test_data, model=scorer)
 print(f"\n{llm}")
 print(f"Number of scores: {len(scores)} of {len(test_data)}")
-print(f"Average score: {sum(scores)/len(scores):.2f}\n")
+print(f"Average score: {sum(scores) / len(scores):.2f}\n")
 ```
 
 This will print out the average score on all test set entries:

tutorials/python-api.md

@@ -104,6 +104,7 @@ To start with random weights, for example, if you plan a pretraining script, ini
 
 ```python
 from litgpt.api import LLM
+
 llm = LLM.load("pythia-160m", init="random", tokenizer_dir="EleutherAI/pythia-160m")
 ```
 
@@ -121,15 +122,12 @@ The `generate_strategy="sequential"` setting loads different parts of the models
 ```python
 from litgpt.api import LLM
 
-llm = LLM.load(
-    "microsoft/phi-2",
-    distribute=None
-)
+llm = LLM.load("microsoft/phi-2", distribute=None)
 
 llm.distribute(
     generate_strategy="sequential",
     devices=4,  # Optional setting, otherwise uses all available GPUs
-    fixed_kv_cache_size=256  # Optionally use a small kv-cache to further reduce memory usage
+    fixed_kv_cache_size=256,  # Optionally use a small kv-cache to further reduce memory usage
 )
 ```
 
@@ -161,11 +159,7 @@ from litgpt.api import LLM
 
 
 if __name__ == "__main__":
-
-    llm = LLM.load(
-        model="meta-llama/Meta-Llama-3.1-8B-Instruct",
-        distribute=None
-    )
+    llm = LLM.load(model="meta-llama/Meta-Llama-3.1-8B-Instruct", distribute=None)
 
     llm.distribute(generate_strategy="tensor_parallel", devices=4)
 
@@ -183,10 +177,7 @@ Use the `.benchmark()` method to compare the computational performance of differ
 from litgpt.api import LLM
 from pprint import pprint
 
-llm = LLM.load(
-    model="microsoft/phi-2",
-    distribute=None
-)
+llm = LLM.load(model="microsoft/phi-2", distribute=None)
 
 llm.distribute(fixed_kv_cache_size=500)
 
@@ -355,7 +346,6 @@ lit_model.llm.generate("hello world")
 The continued pretraining or finetuning from a downloaded model checkpoint is similar to the example above, except that we can skip the initial steps of instantiating a model with random weights.
 
 ```python
-
 lit_model = LitLLM(checkpoint_dir="EleutherAI/pythia-160m")
 data = Alpaca2k()
 
@@ -380,16 +370,16 @@ lit_model.llm.generate("hello world")
 Suppose you trained a model and decide to follow up with a few additional training rounds. This can be achieved as follows by loading an existing Trainer checkpoint:
 
 ```python
-
 import os
 
+
 def find_latest_checkpoint(directory):
     latest_checkpoint = None
     latest_time = 0
 
     for root, _, files in os.walk(directory):
         for file in files:
-            if file.endswith('.ckpt'):
+            if file.endswith(".ckpt"):
                 file_path = os.path.join(root, file)
                 file_time = os.path.getmtime(file_path)
                 if file_time > latest_time:
@@ -398,6 +388,7 @@ def find_latest_checkpoint(directory):
 
     return latest_checkpoint
 
+
 lit_model = LitLLM(checkpoint_dir="EleutherAI/pythia-160m", trainer_ckpt_path=find_latest_checkpoint("lightning_logs"))
 
 data.connect(lit_model.llm.tokenizer, batch_size=batch_size, max_seq_length=512)