Merge branch 'main' into multiline

.github/azure-gpu-test.yml

@@ -41,6 +41,7 @@ jobs:
       displayName: "Image info & NVIDIA"
 
     - script: |
+        pip install --upgrade pip
         pip install '.[all,test]'
       displayName: 'Install dependencies'
 

.github/workflows/check-links.yml

@@ -1,4 +1,4 @@
-name: Check Markdown Links
+name: Check hyperlinks
 
 on:
   push:
@@ -9,30 +9,23 @@ on:
       - main
 
 jobs:
-  check-links:
+  test:
     runs-on: ubuntu-latest
 
     steps:
-      - name: Checkout Repository
-        uses: actions/checkout@v3
+    - uses: actions/checkout@v4
 
-      - name: Install Markdown Link Checker
-        run: npm install -g markdown-link-check
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: '3.10'
 
-      - name: Create config for markdown link checker
-        run: |
-          echo '{
-            "projectBaseUrl":"${{ github.workspace }}",
-            "ignorePatterns": [
-              {
-                "pattern": "^#"
-              },
-              {
-                "pattern": "^https://falconllm.tii.ae"
-              }
-            ]
-          }' > $GITHUB_WORKSPACE/md_checker_config.json
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install pytest pytest-check-links
 
-      - name: Find Markdown Files and Check Links
-        run: |
-          find . -name \*.md -print0 | xargs -0 -n1 markdown-link-check -c $GITHUB_WORKSPACE/md_checker_config.json
+    - name: Check links
+      run: |
+        pytest --check-links README.md --check-links-ignore "https://stability.ai/blog/stablecode-llm-generative-ai-coding"
+        pytest --check-links tutorials --check-links-ignore "https://stability.ai/blog/stablecode-llm-generative-ai-coding"

.github/workflows/cpu-tests.yml

@@ -45,6 +45,7 @@ jobs:
 
     - name: Install minimal dependencies
       run: |
+        pip install --upgrade pip
         pip install .
         pip list
         # make sure all modules are still importable with only the minimal dependencies available

README.md

@@ -5,11 +5,11 @@
 
 # ⚡ LitGPT
 
-**Pretrain, finetune, deploy 20+ LLMs on your own data**
+**Pretrain, finetune, evaluate, and deploy 20+ LLMs on your own data**
 
 Uses the latest state-of-the-art techniques:
 
-✅ fp4/8/16/32      ✅ LoRA, QLoRA, Adapter (v1, v2)      ✅ flash attention      ✅ FSDP      ✅ 1-1000+ GPUs/TPUs
+✅ flash attention      ✅ fp4/8/16/32      ✅ LoRA, QLoRA, Adapter (v1, v2)      ✅ FSDP      ✅ 1-1000+ GPUs/TPUs
 
 ---
 
@@ -32,14 +32,15 @@ Uses the latest state-of-the-art techniques:
 
  
 
-## Finetune, pretrain and deploy AI models Lightning fast ⚡⚡
-LitGPT is a command-line tool to use, pretrain, finetune and deploy LLMs. It is based on configs with highly-optimized recipes for training the world's largest, most powerful open-source LLMs.     
+# Finetune, pretrain and deploy LLMs Lightning fast ⚡⚡   
+LitGPT is a command-line tool designed to easily [finetune](#finetune-an-llm), [pretrain](#pretrain-an-llm), [evaluate](#use-an-llm), and deploy [20+ LLMs](#choose-from-20-llms) **on your own data**. It features highly-optimized [training recipes](#training-recipes) for the world's most powerful open-source large-language-models (LLMs).
 
-We've reimplemented all the model architectures and training recipes for 3 reasons:   
+We reimplemented all model architectures and training recipes from scratch for 4 reasons:   
 
-1. Remove all abstraction layers and have single file implementations.
-2. Guarantee Apache 2.0 compliance to enable enterprise use without limits.
-3. Optimized every detail of every model to get the fastest performance possible to lower cost and training speeds.       
+1. Remove all abstraction layers and have single file implementations.   
+2. Guarantee Apache 2.0 compliance to enable enterprise use without limits.    
+3. Optimized each model architectural detail to maximize performance, reduce costs, and speed up training.    
+4. Highly-optimized [recipe configs](#training-recipes) we have tested at enterprise scale.               
 
  
 
@@ -70,7 +71,16 @@ pip install -e '.[all]'
 ---
 
 # Get started
-LitGPT is CLI and config-based. Select the model and the action you want to take on that model (finetune, pretrain, evaluate, deploy, etc...):    
+After installing LitGPT, select the model and action you want to take on that model (finetune, pretrain, evaluate, deploy, etc...):    
+
+```bash
+# ligpt [action] [model]
+litgpt  download  mistralai/Mistral-7B-Instruct-v0.2
+litgpt  chat      mistralai/Mistral-7B-Instruct-v0.2
+litgpt  finetune  mistralai/Mistral-7B-Instruct-v0.2    
+litgpt  pretrain  mistralai/Mistral-7B-Instruct-v0.2    
+litgpt  serve     mistralai/Mistral-7B-Instruct-v0.2    
+```
 
  
 
@@ -230,9 +240,9 @@ Use, Finetune, pretrain, deploy over 20+ LLMs ([full list](tutorials/download_mo
 
 # Training recipes
 
-LitGPT comes with validated recipes (YAML configs) to train models under different conditions.
+LitGPT comes with validated recipes (YAML configs) to train models under different conditions.  We've generated these recipes based on the parameters we found to perform the best for different training conditions.
 
-We've generated these recipes based on the parameters we found to perform the best for different training conditions.
+Browse all training recipes [here](config_hub).
 
 ### Example
 
@@ -241,8 +251,6 @@ litgpt finetune lora \
   --config https://raw.githubusercontent.com/Lightning-AI/litgpt/main/config_hub/finetune/llama-2-7b/lora.yaml
 ```
 
-Browse all training recipes [here](config_hub).
-
 ### What is a config
 Configs let you customize training for all granular parameters like:
 
@@ -401,6 +409,8 @@ litgpt finetune lora \
 
  
 
+# Community        
+
 ## Get involved!
 
 We appreciate your feedback and contributions. If you have feature requests, questions, or want to contribute code or config files, please don't hesitate to use the [GitHub Issue](https://github.com/Lightning-AI/litgpt/issues) tracker.

litgpt/scripts/convert_lit_checkpoint.py

@@ -235,7 +235,7 @@ def check_conversion_supported(lit_weights: Dict[str, torch.Tensor]) -> None:
     if any("lora" in wn for wn in lit_weights):
         raise ValueError("Checkpoints with LoRA weights cannot be converted. Call `scripts/merge_lora.py` first.")
     if any("adapter" in wn or "gating_factor" in wn for wn in lit_weights):
-        raise NotImplementedError("Converting adapter models is supported.")
+        raise NotImplementedError("Converting adapter models is not supported.")
 
 
 @torch.inference_mode()

pyproject.toml

@@ -23,36 +23,36 @@ litgpt = "litgpt.__main__:main"
 
 [project.optional-dependencies]
 test = [
-    "pytest",
-    "pytest-rerunfailures",
-    "pytest-timeout",
+    "pytest>=8.1.1",
+    "pytest-rerunfailures>=14.0",
+    "pytest-timeout>=2.3.1",
     "transformers>=4.38.0",  # numerical comparisons
-    "einops",
-    "protobuf",
+    "einops>=0.7.0",
+    "protobuf>=4.23.4",
     "lightning-thunder==0.2.0.dev20240404; python_version >= '3.10'",
 ]
 all = [
     "bitsandbytes==0.42.0",      # quantization
-    "sentencepiece",             # llama-based models
-    "tokenizers",                # pythia, falcon, redpajama
-    "requests",                  # litgpt.data
-    "litdata",                   # litgpt.data
-    "zstandard",                 # litgpt.data.prepare_slimpajama.py
-    "pandas",                    # litgpt.data.prepare_starcoder.py
-    "pyarrow",                   # litgpt.data.prepare_starcoder.py
-    "tensorboard",               # litgpt.pretrain
-    "torchmetrics",              # litgpt.pretrain
-    "datasets",                  # litgpt.evaluate
+    "sentencepiece>=0.2.0",      # llama-based models
+    "tokenizers>=0.15.2",        # pythia, falcon, redpajama
+    "requests>=2.31.0",          # litgpt.data
+    "litdata>=0.2.2",            # litgpt.data
+    "zstandard>=0.22.0",         # litgpt.data.prepare_slimpajama.py
+    "pandas>=1.9.0",             # litgpt.data.prepare_starcoder.py
+    "pyarrow>=15.0.2",           # litgpt.data.prepare_starcoder.py
+    "tensorboard>=2.14.0",       # litgpt.pretrain
+    "torchmetrics>=1.3.1",       # litgpt.pretrain
+    "datasets>=2.18.0",          # litgpt.evaluate
     "transformers>=4.38.0",      # litgpt.evaluate
     "lm-eval>=0.4.2",            # litgpt.evaluate
-    "safetensors",               # download
+    "safetensors>=0.4.3",        # download
     "huggingface_hub[hf_transfer]>=0.21.0"  # download
 ]
 
 [build-system]
 requires = [
-    "setuptools",
-    "wheel",
+    "setuptools>=68.2.2",
+    "wheel>=0.41.2",
 ]
 build-backend = "setuptools.build_meta"
 

tests/test_thunder_fsdp.py

@@ -263,6 +263,8 @@ def set_up_planner(self, state_dict, metadata, is_coordinator):
 
 @RunIf(min_cuda_gpus=2, thunder=True, standalone=True)
 def test_save_load_sharded_checkpoint(tmp_path):
+    pytest.skip("Temporarily disabled, often exceeds 5 min timeout")
+
     strategy = ThunderFSDPStrategy(state_dict_type="sharded", broadcast_from=0)
     fabric = Fabric(accelerator="cuda", devices=2, strategy=strategy)
     fabric.launch()

tutorials/download_model_weights.md

@@ -243,7 +243,7 @@ litgpt download \
  
 ## Finetunes and other model variants
 
-Sometimes you want to download the weights of a finetune of one of the models listed above. To do this, you need to manually specifiy the `model_name` associated to the config to use. For example:
+Sometimes you want to download the weights of a finetune of one of the models listed above. To do this, you need to manually specify the `model_name` associated to the config to use. For example:
 
 ```bash
 litgpt download \