Revert "Add core dependency on stable torchdata (#2408) (#2509)"

This reverts commit 4e458d6.

Author: pbontrager

README.md

@@ -11,11 +11,11 @@
 
 ### 📣 Recent updates 📣
 * *February 2025*: Multi-node training is officially [open for business in torchtune](https://pytorch.org/torchtune/main/tutorials/multinode.html)! Full finetune on multiple nodes to take advantage of larger batch sizes and models.
-* *December 2024*: torchtune now supports **Llama 3.3 70B**! Try it out by following our installation instructions [here](#installation-%EF%B8%8F), then run any of the configs [here](recipes/configs/llama3_3).
+* *December 2024*: torchtune now supports **Llama 3.3 70B**! Try it out by following our installation instructions [here](#Installation), then run any of the configs [here](recipes/configs/llama3_3).
 * *November 2024*: torchtune has released [v0.4.0](https://github.com/pytorch/torchtune/releases/tag/v0.4.0) which includes stable support for exciting features like activation offloading and multimodal QLoRA
 * *November 2024*: torchtune has added [Gemma2](recipes/configs/gemma2) to its models!
 * *October 2024*: torchtune added support for Qwen2.5 models - find the configs [here](recipes/configs/qwen2_5/)
-* *September 2024*: torchtune has support for **Llama 3.2 11B Vision**, **Llama 3.2 3B**, and **Llama 3.2 1B** models! Try them out by following our installation instructions [here](#installation-%EF%B8%8F), then run any of the text configs [here](recipes/configs/llama3_2) or vision configs [here](recipes/configs/llama3_2_vision).
+* *September 2024*: torchtune has support for **Llama 3.2 11B Vision**, **Llama 3.2 3B**, and **Llama 3.2 1B** models! Try them out by following our installation instructions [here](#Installation), then run any of the text configs [here](recipes/configs/llama3_2) or vision configs [here](recipes/configs/llama3_2_vision).
 
 
  
@@ -25,9 +25,9 @@
 
 torchtune is a PyTorch library for easily authoring, post-training, and experimenting with LLMs. It provides:
 
-- Hackable training recipes for SFT, knowledge distillation, DPO, PPO, GRPO, and quantization-aware training
+- Hackable training recipes for SFT, knowledge distillation, RL and RLHF, and quantization-aware training
 - Simple PyTorch implementations of popular LLMs like Llama, Gemma, Mistral, Phi, Qwen, and more
-- Best-in-class memory efficiency, performance improvements, and scaling, utilizing the latest PyTorch APIs
+- OOTB best-in-class memory efficiency, performance improvements, and scaling, utilizing the latest PyTorch APIs
 - YAML configs for easily configuring training, evaluation, quantization or inference recipes
 
  

docs/source/api_ref_models.rst

@@ -299,26 +299,6 @@ To download the Qwen2 1.5B model, for example:
     qwen2.lora_qwen2_7b
     qwen2.qwen2_tokenizer
 
-phi-4
------
-
-Models from the `Phi-4 family <https://arxiv.org/abs/2412.08905>`_.
-
-To download the Phi-4 instruct model:
-
-.. code-block:: bash
-
-    tune download microsoft/phi-4 --hf-token <HF_TOKEN>
-
-.. autosummary::
-    :toctree: generated/
-    :nosignatures:
-
-    phi4.phi4_14b
-    phi4.lora_phi4_14b
-    phi4.qlora_phi4_14b
-    phi4.phi4_tokenizer
-
 phi-3
 -----
 

docs/source/api_ref_training.rst

@@ -128,5 +128,4 @@ Miscellaneous
     :nosignatures:
 
     get_unmasked_sequence_lengths
-    disable_dropout
     set_seed

docs/source/deep_dives/checkpointer.rst

@@ -402,8 +402,6 @@ to look for it in the output_dir.
 .. code-block:: yaml
 
     checkpointer:
-        # [... rest of the config...]
-
         # checkpoint files. Note that you will need to update this
         # section of the config with the intermediate checkpoint files
         checkpoint_files: [
@@ -425,11 +423,10 @@ because the base model being loaded is still the same.
 .. code-block:: yaml
 
     checkpointer:
-        # [... rest of the config...]
 
         # adapter_checkpoint. Note that you will need to update this
         # section of the config with the intermediate checkpoint files
-        adapter_checkpoint: epoch_{YOUR_EPOCH}/adapter_model.pt
+        adapter_checkpoint: epoch_{YOUR_EPOCH}/adapter_model.safetensors
 
     # set to True if restarting training
     resume_from_checkpoint: True

docs/source/tutorials/e2e_flow.rst

@@ -29,7 +29,7 @@ Finetune your model
 -------------------
 
 First, let's download a model using the tune CLI. The following command will download the `Llama3.2 3B Instruct <https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/>`_
-model from the Hugging Face Hub and save it to the local filesystem. Hugging Face uploaded the original
+model from the Hugging Face Hub and save it the local filesystem. Hugging Face uploaded the original
 weights (``consolidated.00.pth``) and the weights compatible with the `from_pretrained() <https://huggingface.co/docs/huggingface_hub/main/en/guides/integrations#frompretrained>`_ API (``*.safetensors``).
 We don't need both so we'll ignore the original weights when downloading.
 
@@ -168,15 +168,15 @@ There are 3 types of folders:
 Let's understand the files:
 
 - ``adapter_model.safetensors`` and ``adapter_model.pt`` are your LoRA trained adapter weights. We save a duplicated .pt version of it to facilitate resuming from checkpoint.
-- ``model-{}-of-{}.safetensors`` are your trained full model weights (not adapters). When LoRA finetuning, these are only present if we set ``save_adapter_weights_only=False``. In that case, we merge the base model with trained adapters, making inference easier.
+- ``model-{}-of-{}.safetensors`` are your trained full model weights (not adapters). When LoRA finetuning, these are only present if we set ``save_adapter_weights_only=False``. In that case, we merge the merged base model with trained adapters, making inference easier.
 - ``adapter_config.json`` is used by Huggingface PEFT when loading an adapter (more on that later);
 - ``model.safetensors.index.json`` is used by Hugging Face ``from_pretrained()`` when loading the model weights (more on that later)
-- All other files were originally in the checkpoint_dir. They are automatically copied during training. Files over 100MiB and ending in .safetensors, .pth, .pt, .bin are ignored, making it lightweight.
+- All other files were originally in the checkpoint_dir. They are automatically copied during training. Files over 100MiB and ending on .safetensors, .pth, .pt, .bin are ignored, making it lightweight.
 
 Evaluate your model
 -------------------
 
-We've fine-tuned a model. But how well does this model really do? Let's determine this through structured evaluation and playing with it.
+We've fine-tuned a model. But how well does this model really do? Let's determine this through structured evaluation and playing around with it.
 
 .. _eval_harness_label:
 
@@ -364,8 +364,9 @@ to those in the previously-linked table.
 Use your model in the wild
 --------------------------
 
-Let's say we're happy with how our model is performing at this point - we want to do something with it! Productionize it for serving, publish on the Hugging Face Hub, etc.
-Since we handle checkpoint conversion, you can directly work with standard formats.
+Let's say we're happy with how our model is performing at this point - we want to do something with it! Productionize for serving, publish on the Hugging Face Hub, etc.
+As we mentioned above, one of the benefits of handling of the checkpoint conversion is that you can directly work with standard formats. This helps
+with interoperability with other libraries since torchtune doesn't add yet another format to the mix.
 
 Use with Hugging Face ``from_pretrained()``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -438,8 +439,8 @@ Use with vLLM
 `vLLM <https://docs.vllm.ai/en/latest/>`_ is a fast and easy-to-use library for LLM inference and serving. They include a lot of awesome features like
 state-of-the-art serving throughput, continuous batching of incoming requests, quantization, and speculative decoding.
 
-The library will load any .safetensors file. Since we already merged the full model weights and adapter weights, we can safely delete the
-adapter weights (or move them) so that vLLM doesn't get confused by those files.
+The library will load any .safetensors file. Since here we mixed both the full model weights and adapter weights, we have to delete the
+adapter weights to succesfully load it.
 
 .. code-block:: python
 

pyproject.toml

@@ -10,8 +10,6 @@ authors = [
 ]
 keywords = ["pytorch", "finetuning", "llm"]
 dependencies = [
-    # Stable torchdata (no nightly support)
-    "torchdata",
 
     # Hugging Face integrations
     "datasets",
@@ -35,6 +33,7 @@ dependencies = [
 
     # Multimodal
     "Pillow>=9.4.0",
+
 ]
 dynamic = ["version"]
 

recipes/configs/llama3_2_vision/11B_generation_v2.yaml

@@ -5,7 +5,7 @@
 #  tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct --ignore-patterns "original/consolidated*"
 #
 # To launch, run the following command from root torchtune directory:
-#    tune run dev/generate_v2 --config llama3_2_vision/11B_generation_v2
+#    tune run dev/generate_v2 --config llama3_2_vision/generation_v2
 
 output_dir: ./
 

…/configs/dev/11B_lora_multi_dataset.yaml …ma3_2_vision/11B_lora_multi_dataset.yamlrecipes/configs/dev/11B_lora_multi_dataset.yaml renamed to recipes/configs/llama3_2_vision/11B_lora_multi_dataset.yaml recipes/configs/dev/11B_lora_multi_dataset.yaml renamed to recipes/configs/llama3_2_vision/11B_lora_multi_dataset.yaml

@@ -5,11 +5,11 @@
 #   tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct --ignore-patterns "original/consolidated*"
 #
 # To launch on 2 devices, run the following command from root:
-#   tune run --nproc_per_node 2 dev/lora_finetune_distributed_multi_dataset --config dev/11B_lora_multi_dataset
+#   tune run --nproc_per_node 2 lora_finetune_distributed_multi_dataset --config llama3_2_vision/11B_lora_multi_dataset
 #
 # You can add specific overrides through the command line. For example
 # to override the checkpointer directory while launching training:
-#   tune run --nproc_per_node 2 dev/lora_finetune_distributed_multi_dataset --config dev/11B_lora_multi_dataset checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
+#   tune run --nproc_per_node 2 lora_finetune_distributed_multi_dataset --config llama3_2_vision/11B_lora_multi_dataset checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
 #
 # This config works best when the model is being fine-tuned on 2+ GPUs.
 # For single device LoRA finetuning please use 11B_lora_single_device.yaml
@@ -116,6 +116,7 @@ enable_activation_checkpointing: True
 dtype: bf16
 
 # Logging
+output_dir: /tmp/lora-llama3.2-vision-finetune
 metric_logger:
   _component_: torchtune.training.metric_logging.DiskLogger
   log_dir: ${output_dir}/logs

recipes/configs/mistral/7B_full_ppo_low_memory.yaml

@@ -30,7 +30,7 @@ output_dir: /tmp/torchtune/mistral_7B/full_ppo_low_memory # /tmp may be deleted
 tokenizer:
   _component_: torchtune.models.mistral.mistral_tokenizer
   path: /tmp/Mistral-7B-Instruct-v0.2/tokenizer.model
-  max_seq_len: 512
+  max_seq_len: null
 
 # Dataset
 dataset:

recipes/configs/phi4/evaluation.yaml

@@ -14,21 +14,17 @@ checkpointer:
   _component_: torchtune.training.FullModelHFCheckpointer
   checkpoint_dir: /tmp/phi-4
   checkpoint_files: [
-    model-00001-of-00006.safetensors,
-    model-00002-of-00006.safetensors,
-    model-00003-of-00006.safetensors,
-    model-00004-of-00006.safetensors,
-    model-00005-of-00006.safetensors,
-    model-00006-of-00006.safetensors,
+    model-00001-of-00002.safetensors,
+    model-00002-of-00002.safetensors
   ]
   recipe_checkpoint: null
   output_dir: ${output_dir}
-  model_type: PHI4
+  model_type: PHI3_MINI
 resume_from_checkpoint: False
 
 # Tokenizer
 tokenizer:
-  _component_: torchtune.models.phi4.phi4_tokenizer
+  _component_: torchtune.models.phi4.phi4_14b_tokenizer
   vocab_path: /tmp/phi-4/vocab.json
   merges_path: /tmp/phi-4/merges.txt
   max_seq_len: null