Revert "checkpoint model on first step callback (#2906)"

This reverts commit 10ba1622f7.

.github/workflows/multi-gpu-e2e.yml

@@ -33,6 +33,13 @@ jobs:
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"

.github/workflows/nightlies.yml

@@ -12,11 +12,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -60,15 +65,15 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:

.github/workflows/preview-docs.yml

@@ -28,6 +28,8 @@ jobs:
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
 
       - name: Set up Quarto
         uses: quarto-dev/quarto-actions/setup@v2
@@ -50,10 +52,11 @@ jobs:
 
       - name: Netlify Publish
         uses: nwtgck/actions-netlify@v3.0
+        id: netlify
         with:
           publish-dir: './_site'
-          enable-pull-request-comment: true
-          enable-github-deployment: true
+          enable-pull-request-comment: false
+          enable-github-deployment: false
           github-token: ${{ secrets.GITHUB_TOKEN }}
           deploy-message: "Deployed On Netlify"
           github-deployment-environment: 'preview'
@@ -61,3 +64,13 @@ jobs:
         env:
           NETLIFY_AUTH_TOKEN: ${{ secrets.NETLIFY_AUTH_TOKEN }}
           NETLIFY_SITE_ID: ${{ secrets.NETLIFY_SITE_ID }}
+
+      - name: Update PR with preview link
+        if: ${{ steps.netlify.outcome == 'success' }}
+        uses: marocchino/sticky-pull-request-comment@v2
+        with:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          message: |
+            📖 **Documentation Preview**: ${{ steps.netlify.outputs.deploy-url }}
+
+            Deployed on Netlify from commit ${{ github.event.pull_request.head.sha }}

.runpod/src/config/config.yaml

@@ -97,7 +97,7 @@
 #       # 'no_input_format' cannot include {input}
 #       no_input_format: "{instruction} "
 
-#       # For `completion` datsets only, uses the provided field instead of `text` column
+#       # For `completion` datasets only, uses the provided field instead of `text` column
 #       field:
 
 # # Axolotl attempts to save the dataset as an arrow after packing the data together so

_quarto.yml

@@ -276,6 +276,7 @@ website:
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
+            - docs/gradient_checkpointing.qmd
 
         - section: "Troubleshooting"
           contents:

docs/dataset-formats/conversation.qmd

@@ -187,6 +187,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
             "role": "assistant", // call the function via assistant
             "tool_calls": [
                 {
+                    "id": "...",  // required only for mistral
                     "type": "function",
                     "function": {
                         "name": "...",
@@ -199,6 +200,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
         },
         {
             "role": "tool",
+            "tool_call_id": "...",  // required only for mistral
             "name": "...",
             "content": "..."
         },

docs/docker.qmd

@@ -34,6 +34,7 @@ Tags examples:
 
 - `main-base-py3.11-cu128-2.7.1`
 - `main-base-py3.11-cu126-2.7.1`
+- `main-base-py3.11-cu126-2.7.0`
 - `main-base-py3.11-cu126-2.6.0`
 - `main-base-py3.11-cu124-2.6.0`
 
@@ -75,6 +76,7 @@ Tags examples:
 
 - `main-py3.11-cu128-2.7.1`
 - `main-py3.11-cu126-2.7.1`
+- `main-py3.11-cu126-2.7.0`
 - `main-py3.11-cu126-2.6.0`
 - `main-py3.11-cu124-2.6.0`
 - `main-latest`

docs/gradient_checkpointing.qmd

@@ -0,0 +1,29 @@
+---
+title: Gradient Checkpointing and Activation Offloading
+---
+
+Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
+models by reducing the memory footprint and improving computational efficiency.
+
+### Enabling Gradient Checkpointing
+
+```yaml
+gradient_checkpointing: true
+```
+
+### Enabling Activation Offloading
+
+```yaml
+gradient_checkpointing: true  # required for activation offloading
+activation_offloading: true
+```
+
+Activation offloading variants:
+
+The default `activation_offloading: true` offloads activations to CPU and uses CUDA streams
+to overlap the communications and computations when offloading.
+
+The `activation_offloading: legacy` naively offloads activations to CPU and without additional optimizations.
+
+For resource constrained environments with limited CPU memory, `activation_offloading: disk` offloads
+activations to disk instead of CPU RAM so that much larger context lengths can be trained with minimal memory.

docs/multi-gpu.qmd

@@ -23,8 +23,6 @@ Axolotl supports several methods for multi-GPU training:
 
 ## DeepSpeed {#sec-deepspeed}
 
-DeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.
-
 ### Configuration {#sec-deepspeed-config}
 
 Add to your YAML config:
@@ -32,7 +30,6 @@ Add to your YAML config:
 ```{.yaml}
 deepspeed: deepspeed_configs/zero1.json
 ```
-
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
@@ -66,9 +63,75 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-## FSDP {#sec-fsdp}
+::: {.callout-tip}
 
-### Basic FSDP Configuration {#sec-fsdp-config}
+Using ZeRO Stage 3 with Single-GPU training
+
+ZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:
+`WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500`
+
+:::
+
+## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
+
+::: {.callout-note}
+
+FSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.
+
+:::
+
+### Migrating from FSDP1 to FSDP2 {#sec-migrate-fsdp1-fsdp2}
+
+To migrate your config from FSDP1 to FSDP2, you must use the `fsdp_version` top-level config field to specify the FSDP version, and
+also follow the config field mapping below to update field names.
+
+#### Config mapping
+
+FSDP1 | FSDP2
+-------- | --------
+fsdp_sharding_strategy | reshard_after_forward
+fsdp_backward_prefetch_policy | **REMOVED**
+fsdp_backward_prefetch | **REMOVED**
+fsdp_forward_prefetch | **REMOVED**
+fsdp_sync_module_states | **REMOVED**
+fsdp_cpu_ram_efficient_loading | cpu_ram_efficient_loading
+fsdp_state_dict_type | state_dict_type
+fsdp_use_orig_params | **REMOVED**
+
+
+For example, if you were using the following FSDP1 config:
+
+```{.yaml}
+fsdp_version: 1
+fsdp_config:
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+```
+
+You can migrate to the following FSDP2 config:
+
+```{.yaml}
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+```
+
+### FSDP1 (deprecated) {#sec-fsdp-config}
+
+::: {.callout-note}
+
+Using `fsdp` to configure FSDP is deprecated and will be removed in an upcoming release of Axolotl. Please use `fsdp_config` as above instead.
+
+:::
 
 ```{.yaml}
 fsdp:
@@ -80,6 +143,7 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+
 ## Sequence parallelism {#sec-sequence-parallelism}
 
 We support sequence parallelism (SP) via the

docs/multi-node.qmd

@@ -40,13 +40,13 @@ use_cpu: false
 
 Configure your model to use FSDP in the Axolotl yaml. For example:
 ```yaml
-fsdp:
-  - full_shard
-  - auto_wrap
+fsdp_version: 2
 fsdp_config:
-  fsdp_offload_params: true
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  offload_params: true
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
 ```
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.

docs/rlhf.qmd

@@ -17,7 +17,6 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl, if you're interested in contributing, please reach out!)
 
 
 ## RLHF using Axolotl
@@ -275,15 +274,14 @@ rl: dpo
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_chosen: "chosen"
-    field_rejected: "rejected"
-    prompt_format: "{prompt}"
-    chosen_format: "{chosen}"
-    rejected_format: "{rejected}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_chosen: "chosen"
+      field_rejected: "rejected"
+      prompt_format: "{prompt}"
+      chosen_format: "{chosen}"
+      rejected_format: "{rejected}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.
@@ -476,14 +474,13 @@ rl: kto
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_completion: "completion"
-    field_label: "label"
-    prompt_format: "{prompt}"
-    completion_format: "{completion}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_completion: "completion"
+      field_label: "label"
+      prompt_format: "{prompt}"
+      completion_format: "{completion}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.

examples/archived/README.md

@@ -0,0 +1,5 @@
+# Archived Examples
+
+This directory contains examples that are no longer maintained and may no longer be functional.
+
+We keep them around for archival purposes in case they are useful to others.

examples/devstral/README.md

@@ -1,8 +1,12 @@
 # Finetune Devstral with Axolotl
 
-Devstral Small is a 24B parameter opensource model from MistralAI found on HuggingFace [Devstral-Small-2505](https://huggingface.co/mistralai/Devstral-Small-2505). This guide shows how to fine-tune it with Axolotl with multi-turn conversations with proper masking.
+Devstral Small is a 24B parameter opensource model from MistralAI found on HuggingFace [Devstral-Small-2505](https://huggingface.co/mistralai/Devstral-Small-2505) and [Devstral-Small-2507](https://huggingface.co/mistralai/Devstral-Small-2507). `Devstral-Small-2507` is the latest version of the model and has [function calling](https://mistralai.github.io/mistral-common/usage/tools/) support.
 
-The model was fine-tuned ontop of [Mistral-Small-3.1](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Base-2503) without the vision layer and has a context of upto 128k tokens.
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations with proper masking.
+
+The model was fine-tuned ontop of [Mistral-Small-3.1](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Base-2503) without the vision layer and has a context of up to 128k tokens.
+
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
 
 ## Getting started
 
@@ -17,11 +21,6 @@ cd axolotl
 
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
-
-# Install the latest mistral-common from source
-pip3 uninstall mistral-common
-pip3 install git+https://github.com/mistralai/mistral-common.git@039465d
-
 ```
 
 2. Run the finetuning example:
@@ -39,6 +38,7 @@ Let us know how it goes. Happy finetuning! 🚀
 - You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
 - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
 - The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- Learn how to use function calling with Axolotl at [docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use).
 
 ## Optimization Guides
 
@@ -57,6 +57,7 @@ In addition, we do not support overriding tokens yet.
 ## Related Resources
 
 - [MistralAI Devstral Blog](https://mistral.ai/news/devstral)
+- [MistralAI Devstral 1.1 Blog](https://mistral.ai/news/devstral-2507)
 - [Axolotl Docs](https://docs.axolotl.ai)
 - [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
 - [Axolotl Website](https://axolotl.ai)

examples/devstral/devstral-small-qlora.yml

@@ -1,4 +1,4 @@
-base_model: mistralai/Devstral-Small-2505
+base_model: mistralai/Devstral-Small-2507
 
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name

examples/lfm2/README.md

@@ -0,0 +1,7 @@
+# Liquid Foundation Models 2
+
+LFM2 support in transformers exists in the main branch, but is not yet included in the transformers release.
+
+```bash
+pip install --upgrade --no-deps --force-reinstall git+https://github.com/huggingface/transformers.git
+```

examples/lfm2/lfm2-350m-fft.yaml

@@ -0,0 +1,48 @@
+base_model: LiquidAI/LFM2-350M
+
+chunked_cross_entropy: true
+
+chat_template: tokenizer_default
+eot_tokens:
+  - "<|im_end|>"
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 4
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+
+weight_decay: 0.0

requirements.txt

@@ -6,19 +6,19 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.10
+liger-kernel==0.6.0
 # END section
 
 packaging==23.2
 
-huggingface_hub==0.32.2
-peft==0.15.2
-transformers==4.53.1
+huggingface_hub>=0.33.0
+peft==0.16.0
+transformers==4.53.2
 tokenizers>=0.21.1
 accelerate==1.8.1
-datasets==3.6.0
+datasets==4.0.0
 deepspeed>=0.17.0
-trl==0.18.2
+trl==0.19.1
 hf_xet==1.1.2
 
 optimum==1.16.2
@@ -68,4 +68,4 @@ schedulefree==1.4.1
 axolotl-contribs-lgpl==0.0.6
 axolotl-contribs-mit==0.0.3
 
-mistral-common==1.6.3
+mistral-common==1.7.0

setup.py

@@ -73,9 +73,9 @@ def parse_requirements(extras_require_map):
                 extras_require_map["vllm"] = ["vllm>=0.9.0"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append(
-                    "xformers==0.0.29.post2"
-                )  # vllm needs post2 w torch 2.6
+                _install_requires.append("xformers==0.0.29.post3")
+                # since we only support 2.6.0+cu126
+                _dependency_links.append("https://download.pytorch.org/whl/cu126")
                 extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
@@ -121,7 +121,7 @@ extras_require = {
         "yunchang==0.6.0",
     ],
     "deepspeed": [
-        "deepspeed==0.17.1",
+        "deepspeed==0.17.2",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.11.0"
+__version__ = "0.12.0.dev"

src/axolotl/cli/preprocess.py

@@ -1,5 +1,6 @@
 """CLI to run preprocessing of a dataset."""
 
+import os
 import warnings
 from pathlib import Path
 from typing import Union
@@ -95,6 +96,7 @@ def do_cli(
         kwargs: Additional keyword arguments to override config file values.
     """
     # pylint: disable=duplicate-code
+    os.environ["AXOLOTL_IS_PREPROCESS"] = "1"
     parsed_cfg = load_cfg(config, **kwargs)
     parsed_cfg.is_preprocess = True
     parser = transformers.HfArgumentParser(PreprocessCliArgs)

src/axolotl/cli/train.py

@@ -109,6 +109,13 @@ def ray_train_func(kwargs: dict):
     # initialize accelerator before model instantiation
     Accelerator(gradient_accumulation_steps=cfg.gradient_accumulation_steps)
 
+    # Register plugins in Ray workers
+    if cfg.get("plugins"):
+        from axolotl.cli.config import plugin_set_cfg, prepare_plugins
+
+        prepare_plugins(cfg)
+        plugin_set_cfg(cfg)
+
     kwargs["cfg"] = cfg
 
     do_train(**kwargs)

src/axolotl/cli/vllm_serve.py

@@ -37,7 +37,6 @@ def do_vllm_serve(
     Returns:
         process_id: the process id of the started VLLM server
     """
-    patch_vllm_worker()
     cfg = load_cfg(config)
     model = cfg.base_model
 
@@ -47,6 +46,9 @@ def do_vllm_serve(
     tensor_parallel_size = (
         cli_args.get("tensor_parallel_size") or cfg.vllm.tensor_parallel_size
     )
+    data_parallel_size = (
+        cli_args.get("data_parallel_size") or cfg.vllm.data_parallel_size
+    )
     host = cli_args.get("host") or cfg.vllm.host
     port = cli_args.get("port") or cfg.vllm.port
     gpu_memory_utilization = (
@@ -68,6 +70,7 @@ def do_vllm_serve(
     vllm_script_args = AxolotlScriptArguments(
         model=model,
         tensor_parallel_size=tensor_parallel_size,
+        data_parallel_size=data_parallel_size,
         host=host,
         port=port,
         gpu_memory_utilization=gpu_memory_utilization,

src/axolotl/core/builders/base.py

@@ -112,13 +112,6 @@ class TrainerBuilderBase(abc.ABC):
             plugin_manager.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
         )
 
-        if self.cfg.profiler_steps:
-            callbacks.append(
-                PytorchProfilerCallback(
-                    steps_to_profile=self.cfg.profiler_steps,
-                )
-            )
-
         if self.cfg.gc_steps:
             callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
 
@@ -145,6 +138,14 @@ class TrainerBuilderBase(abc.ABC):
 
         callbacks.append(GPUStatsCallback(cfg=self.cfg))
 
+        if self.cfg.profiler_steps:
+            callbacks.append(
+                PytorchProfilerCallback(
+                    steps_to_profile=self.cfg.profiler_steps,
+                    profiler_steps_start=self.cfg.profiler_steps_start,
+                )
+            )
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -418,6 +419,9 @@ class TrainerBuilderBase(abc.ABC):
             torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
                 True
             )
+            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
+                256
+            )
             training_args_kwargs["torch_compile"] = self.cfg.torch_compile
             if self.cfg.torch_compile_backend:
                 training_args_kwargs["torch_compile_backend"] = (
@@ -426,8 +430,16 @@ class TrainerBuilderBase(abc.ABC):
             if self.cfg.torch_compile_mode:
                 training_args_kwargs["torch_compile_mode"] = self.cfg.torch_compile_mode
 
+    def _configure_accelerator_config(self, training_args_kwargs: dict):
+        if self.cfg.accelerator_config:
+            training_args_kwargs["accelerator_config"] = self.cfg.accelerator_config
+
     def _configure_gradient_checkpointing(self, training_args_kwargs: dict):
-        if self.cfg.gradient_checkpointing:
+        if self.cfg.activation_offloading is True:
+            # don't use the HF gradient checkpointing, manually wrap
+            training_args_kwargs["gradient_checkpointing"] = False
+            training_args_kwargs["activation_offloading"] = True
+        elif self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
             )
@@ -501,10 +513,15 @@ class TrainerBuilderBase(abc.ABC):
         if self.cfg.reward_model or self.cfg.rl:
             training_args_kwargs["max_length"] = self.cfg.sequence_len
 
+        if self.cfg.fsdp_config or self.cfg.fsdp:
+            training_args_kwargs["fsdp_config"] = self.cfg.fsdp_config
+            training_args_kwargs["fsdp"] = self.cfg.fsdp if self.cfg.fsdp else True
+
         self._configure_reporting(training_args_kwargs)
         self._configure_hub_parameters(training_args_kwargs)
         self._configure_scheduler(training_args_kwargs)
         self._configure_optimizer(training_args_kwargs, trainer_kwargs)
         self._configure_torch_compile(training_args_kwargs)
+        self._configure_accelerator_config(training_args_kwargs)
 
         return training_args_kwargs, trainer_kwargs

src/axolotl/core/builders/causal.py

@@ -151,14 +151,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs, trainer_kwargs = self._set_base_training_args(
             total_num_steps
         )
-
-        if self.cfg.fsdp:
-            training_arguments_kwargs["fsdp"] = self.cfg.fsdp
-            if self.cfg.fsdp_config:
-                training_arguments_kwargs["fsdp_config"] = {
-                    k.lstrip("fsdp_"): v for k, v in dict(self.cfg.fsdp_config).items()
-                }
-
         if self.cfg.adapter == "qlora":
             training_arguments_kwargs["qlora"] = True
 
@@ -318,11 +310,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.neftune_noise_alpha
             )
 
-        if self.cfg.accelerator_config:
-            training_arguments_kwargs["accelerator_config"] = (
-                self.cfg.accelerator_config
-            )
-
         if self.cfg.image_size:
             training_arguments_kwargs["image_size"] = self.cfg.image_size
         if self.cfg.image_resize_algorithm:

src/axolotl/core/builders/rl.py

@@ -208,7 +208,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             callbacks=self.get_callbacks(),
             **trainer_kwargs,
         )
-        if self.cfg.fsdp:
+        if self.cfg.fsdp_config or self.cfg.fsdp:
             ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
             if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
                 ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
@@ -218,21 +218,3 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer.add_callback(callback)
 
         return trainer
-
-
-class HFPPOTrainerBuilder(TrainerBuilderBase):
-    """
-    HF Factory class for PPO Trainer
-    """
-
-    def get_callbacks(self):
-        callbacks = super().get_callbacks()
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
-        return callbacks
-
-    def build(self, total_num_steps):
-        # TODO: build PPOConfig
-        raise NotImplementedError("PPO trainer builder is not implemented yet.")

src/axolotl/core/trainers/__init__.py

@@ -14,5 +14,4 @@ from .trl import (
     AxolotlORPOTrainer,
     AxolotlPRMTrainer,
     AxolotlRewardTrainer,
-    TRLPPOTrainer,
 )

src/axolotl/core/trainers/base.py

@@ -25,6 +25,7 @@ from trl.trainer.utils import pad_to_length
 from typing_extensions import override
 
 from axolotl.core.trainers.mixins import (
+    ActivationOffloadingMixin,
     CheckpointSaveMixin,
     OptimizerMixin,
     PackingMixin,
@@ -48,6 +49,7 @@ class AxolotlTrainer(
     OptimizerMixin,
     RngLoaderMixin,
     CheckpointSaveMixin,
+    ActivationOffloadingMixin,
     Trainer,
 ):
     """Extend the base Trainer for axolotl helpers"""
@@ -75,18 +77,6 @@ class AxolotlTrainer(
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.torch_compile:
-            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
-                256
-            )
-            model = torch.compile(
-                model,
-                backend=self.args.torch_compile_backend,
-                mode=self.args.torch_compile_mode,
-            )
-        return super()._wrap_model(model, training=training, dataloader=dataloader)
-
     def _create_multipack_sampler(
         self, base_sampler: Sampler, dataset: Dataset
     ) -> MultipackBatchSampler:

src/axolotl/core/trainers/grpo/__init__.py

@@ -14,6 +14,7 @@ from axolotl.core.trainers.grpo.trainer import (
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.trl import TRLConfig
+from axolotl.utils.schemas.vllm import VllmConfig
 
 LOG = get_logger(__name__)
 
@@ -41,9 +42,18 @@ class GRPOStrategy:
             return grpo_args_kwargs
 
         trl: TRLConfig = cfg.trl  # type: ignore
+        vllm_cfg: VllmConfig = cfg.vllm  # type: ignore
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
+            grpo_args_kwargs["vllm_mode"] = trl.vllm_mode
+            if trl.vllm_mode == "colocate":
+                grpo_args_kwargs["vllm_gpu_memory_utilization"] = (
+                    vllm_cfg.gpu_memory_utilization
+                )
+                grpo_args_kwargs["vllm_tensor_parallel_size"] = (
+                    vllm_cfg.tensor_parallel_size
+                )
             grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host  # type: ignore[attr-defined]
             grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port  # type: ignore[attr-defined]
             if trl.vllm_server_timeout:

src/axolotl/core/trainers/grpo/trainer.py

@@ -59,42 +59,6 @@ class AxolotlGRPOTrainer(
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-    def get_train_dataloader(self):
-        if self.train_dataset is None:
-            raise ValueError("Trainer: training requires a train_dataset.")
-
-        train_dataset = self.train_dataset
-        data_collator = self.data_collator
-        if isinstance(train_dataset, datasets.Dataset):
-            train_dataset = self._remove_unused_columns(
-                train_dataset, description="training"
-            )
-        else:
-            data_collator = self._get_collator_with_removed_columns(
-                data_collator, description="training"
-            )
-
-        dataloader_params = {
-            "batch_size": self._train_batch_size
-            * self.args.steps_per_generation,  # < this is the change
-            "collate_fn": data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-            "persistent_workers": self.args.dataloader_persistent_workers,
-        }
-
-        if not isinstance(train_dataset, torch.utils.data.IterableDataset):
-            dataloader_params["sampler"] = self._get_train_sampler()
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
-            dataloader_params["worker_init_fn"] = partial(
-                seed_worker,
-                num_workers=self.args.dataloader_num_workers,
-                rank=self.args.process_index,
-            )
-            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        return self.accelerator.prepare(DataLoader(train_dataset, **dataloader_params))
-
 
 class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
     """Extend the base GRPOTrainer for sequence parallelism handling"""
@@ -252,7 +216,11 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
                 dataloader_params["drop_last"] = self.args.dataloader_drop_last
 
             if not is_eval:
-                dataloader_params["worker_init_fn"] = seed_worker
+                dataloader_params["worker_init_fn"] = partial(
+                    seed_worker,
+                    num_workers=self.args.dataloader_num_workers,
+                    rank=self.args.process_index,
+                )
 
         # Create the dataloader
         dataloader = DataLoader(dataset, **dataloader_params)

src/axolotl/core/trainers/mixins/__init__.py

@@ -3,6 +3,7 @@
 # pylint: disable=unused-import
 # flake8: noqa
 
+from .activation_checkpointing import ActivationOffloadingMixin
 from .checkpoints import CheckpointSaveMixin
 from .optimizer import OptimizerMixin
 from .packing import PackingMixin

src/axolotl/core/trainers/mixins/activation_checkpointing.py

@@ -0,0 +1,37 @@
+"""
+Trainer mixin for activation checkpointing w offloading
+"""
+
+import contextlib
+
+from torch import nn
+from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
+    apply_activation_checkpointing,
+)
+from torch.distributed.fsdp.wrap import ModuleWrapPolicy
+from transformers import GradientCheckpointingLayer, Trainer
+from trl.models.activation_offloading import get_act_offloading_ctx_manager
+
+
+class ActivationOffloadingMixin(Trainer):
+    """
+    Trainer mixin class for activation checkpointing w offloading
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.args.activation_offloading:
+            self.activation_offload_context = get_act_offloading_ctx_manager(
+                self.model, use_streams=True
+            )
+        else:
+            self.activation_offload_context = contextlib.nullcontext()
+
+    def training_step(self, *args, **kwargs):
+        with self.activation_offload_context:
+            return super().training_step(*args, **kwargs)
+
+
+def ac_wrap_hf_model(model: nn.Module, **kwargs):
+    auto_wrap_policy = ModuleWrapPolicy(set((GradientCheckpointingLayer,)))
+    apply_activation_checkpointing(model, auto_wrap_policy=auto_wrap_policy, **kwargs)

src/axolotl/core/trainers/trl.py

@@ -1,12 +1,9 @@
-"""Module for TRL PPO trainer"""
+"""Module for TRL RL trainers"""
 
-import torch
-from tqdm import tqdm
 from trl import (
     CPOTrainer,
     KTOTrainer,
     ORPOTrainer,
-    PPOTrainer,
     PRMTrainer,
     RewardTrainer,
 )
@@ -16,64 +13,6 @@ from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, Optimizer
 from axolotl.core.trainers.mixins.scheduler import SchedulerMixin
 
 
-class TRLPPOTrainer(PPOTrainer):
-    """Wrapper for TRL PPO trainer to handle customizations"""
-
-    tag_names = ["axolotl", "ppo"]
-
-    def train(
-        self,
-        reward_pipe,
-        resume_from_checkpoint=None,  # pylint: disable=unused-argument
-    ):
-        generation_kwargs = {
-            "min_length": -1,
-            "top_k": 0.0,
-            "top_p": 1.0,
-            "do_sample": True,
-            "pad_token_id": self.tokenizer.eos_token_id,
-            "max_new_tokens": 32,
-        }
-        sent_kwargs = {
-            "return_all_scores": True,
-            "function_to_apply": "none",
-            "batch_size": 16,
-        }
-
-        for _, batch in tqdm(enumerate(self.dataloader)):
-            query_tensors = batch["input_ids"]
-
-            # generate model response
-            response_tensors, ref_response_tensors = self.generate(
-                query_tensors,
-                return_prompt=False,
-                generate_ref_response=True,
-                **generation_kwargs,
-            )
-            batch["response"] = self.tokenizer.batch_decode(response_tensors)
-            batch["ref_response"] = self.tokenizer.batch_decode(ref_response_tensors)
-
-            # Compute sentiment score
-            texts = [q + r for q, r in zip(batch["query"], batch["response"])]
-            pipe_outputs = reward_pipe(texts, **sent_kwargs)
-            rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
-            ref_texts = [q + r for q, r in zip(batch["query"], batch["ref_response"])]
-            ref_pipe_outputs = reward_pipe(ref_texts, **sent_kwargs)
-            ref_rewards = [
-                torch.tensor(output[1]["score"]) for output in ref_pipe_outputs
-            ]
-            batch["ref_rewards"] = ref_rewards
-
-            # Run PPO step
-            stats = self.step(query_tensors, response_tensors, rewards)
-            self.log_stats(
-                stats,
-                batch,
-                rewards,
-                columns_to_log=["query", "response", "ref_response", "ref_rewards"],
-            )
-
-
 class AxolotlORPOTrainer(
     RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, ORPOTrainer
 ):

src/axolotl/core/training_args_base.py

@@ -217,6 +217,11 @@ class AxolotlTrainingMixins:
         },
     )
 
+    activation_offloading: bool | None = field(
+        default=None,
+        metadata={"help": "Use activation offloading with CUDA streams for training."},
+    )
+
     # multi-modal section
 
     image_size: int | tuple[int, int] | None = field(

src/axolotl/integrations/kd/kernels/models.py

@@ -6,15 +6,21 @@ from typing import Optional, Union, Unpack
 
 import torch
 from transformers import Cache
-from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils import LossKwargs
 
+try:
+    from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+    from transformers.utils import LossKwargs
 
-class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs):
-    """
-    placeholder kwargs for hf model classes
-    """
+    class TransformersKwargs(FlashAttentionKwargs, LossKwargs):
+        """
+        placeholder kwargs for hf model classes
+        """
+
+except ImportError:
+    from transformers.utils.generic import (  # type: ignore[no-redef]
+        TransformersKwargs,
+    )
 
 
 def kldiv_forward_llama_like(
@@ -33,7 +39,7 @@ def kldiv_forward_llama_like(
     output_hidden_states: Optional[bool] = None,
     cache_position: Optional[torch.LongTensor] = None,
     logits_to_keep: Union[int, torch.Tensor] = 0,  # pylint: disable=unused-argument
-    **kwargs: Unpack[KwargsForCausalLM],  # type: ignore[misc]
+    **kwargs: Unpack[TransformersKwargs],  # type: ignore[misc]
 ) -> CausalLMOutputWithPast:
     # pylint: disable=duplicate-code
     output_attentions = (

src/axolotl/loaders/adapter.py

@@ -122,9 +122,9 @@ def load_lora(
     rank = int(os.environ.get("LOCAL_RANK", 0))
 
     if (
-        cfg.fsdp
+        cfg.fsdp_config
         and cfg.adapter
-        and cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+        and cfg.fsdp_config.cpu_ram_efficient_loading
         and rank != 0
     ):
         setup_quantized_meta_for_peft(model)
@@ -152,9 +152,9 @@ def load_lora(
                 "Exception caught during model.print_trainable_parameters(): %s", exc
             )
     elif (
-        cfg.fsdp
+        cfg.fsdp_config
         and cfg.adapter
-        and cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+        and cfg.fsdp_config.cpu_ram_efficient_loading
         and rank != 0
     ):
         setup_quantized_peft_meta_for_training(model)

src/axolotl/loaders/model.py

@@ -140,10 +140,15 @@ class ModelLoader:
         """Check if flash attention is installed."""
         return find_spec("flash_attn") is not None
 
-    @cached_property
-    def qlora_fsdp(self):
+    @property
+    def is_fsdp_enabled(self):
+        """Property that determines if FSDP is enabled."""
+        return self.cfg.fsdp_config is not None or self.cfg.fsdp is not None
+
+    @property
+    def is_qlora_and_fsdp_enabled(self):
         """Property that determines if FSDP with QLoRA is enabled."""
-        return self.cfg.fsdp and self.cfg.adapter == "qlora"
+        return self.is_fsdp_enabled and self.cfg.adapter == "qlora"
 
     def load(self) -> tuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]:
         """Load and prepare the model with all configurations and patches.
@@ -189,15 +194,25 @@ class ModelLoader:
         # Handle PeftModel if needed
         if (
             isinstance(self.model, (peft.PeftModel, peft.PeftModelForCausalLM))
-            and not self.qlora_fsdp
+            and not self.is_qlora_and_fsdp_enabled
         ):
             self.model = self.model.merge_and_unload()
 
+        self._apply_activation_checkpointing()
         self._resize_token_embeddings()
         self._adjust_model_config()
-        self._log_memory_usage()
         self._configure_embedding_dtypes()
         self._configure_qat()
+        log_gpu_memory_usage(LOG, "Memory usage after model load", 0)
+
+    def _apply_activation_checkpointing(self):
+        if self.cfg.activation_offloading is True:
+            from axolotl.core.trainers.mixins.activation_checkpointing import (
+                ac_wrap_hf_model,
+            )
+
+            # ^^ importing this at the module level breaks plugins
+            ac_wrap_hf_model(self.model)
 
     def _resize_token_embeddings(self):
         """Resize token embeddings if needed."""
@@ -251,22 +266,13 @@ class ModelLoader:
         ):
             self.model.config.eos_token_id = self.tokenizer.eos_token_id
 
-    def _log_memory_usage(self):
-        """Log device memory usage after model load."""
-        if hasattr(self.model, "device") and self.model.device.type in (
-            "cuda",
-            "mps",
-            "npu",
-        ):
-            log_gpu_memory_usage(LOG, "after model load", self.model.device)
-
     def _configure_embedding_dtypes(self):
         """Configure embedding module dtypes."""
         # Get embedding modules
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
 
         # Initial dtype conversion
-        if not self.cfg.fsdp:
+        if not self.is_fsdp_enabled:
             # We don't run this during FSDP because this will leave mixed and bfloat16
             # dtypes in the model which FSDP doesn't like
             if self.cfg.load_in_4bit and self.cfg.embeddings_skip_upcast:
@@ -282,7 +288,7 @@ class ModelLoader:
             self._set_z3_leaf_modules()
 
         # Apply gradient checkpointing if needed
-        needs_fa2_dtype = self.cfg.adapter or self.cfg.fsdp
+        needs_fa2_dtype = self.cfg.adapter or self.is_fsdp_enabled
         if self.cfg.adapter in ["lora", "qlora"]:
             needs_fa2_dtype = True
             if self.cfg.gradient_checkpointing:
@@ -298,10 +304,12 @@ class ModelLoader:
             # we need to convert them back to fp16/bf16 for flash-attn compatibility.
             (
                 (needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention)
-                and not self.qlora_fsdp
+                and not self.is_qlora_and_fsdp_enabled
+            )
+            or (
+                # CCE requires embedding layers to be in fp16/bf16 for backward pass
+                self.cfg.cut_cross_entropy
             )
-            # CCE requires embedding layers to be in fp16/bf16 for backward pass
-            or self.cfg.cut_cross_entropy
         )
 
         if should_convert:
@@ -357,7 +365,6 @@ class ModelLoader:
             and not (self.cfg.rl and self.cfg.load_in_4bit)
             and not skip_move_to_device
         ):
-            # TODO: validate this conditional
             self.model.to(f"{str(get_device_type())}:{self.cfg.local_rank}")
 
         if get_device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
@@ -430,7 +437,17 @@ class ModelLoader:
 
         self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
 
-        if not is_deepspeed_zero3_enabled():
+        is_ds_zero3 = is_deepspeed_zero3_enabled()
+
+        # FSDP requires control over device placement, so don't set device_map when FSDP is enabled
+        if self.is_fsdp_enabled:
+            # For QLoRA + FSDP, we still need to set device_map to "auto" for proper initialization
+            if self.is_qlora_and_fsdp_enabled:
+                self.model_kwargs["device_map"] = {
+                    "": int(os.environ.get("LOCAL_RANK", 0))
+                }
+            # For other FSDP cases, don't set device_map at all
+        elif not is_ds_zero3:
             self.model_kwargs["device_map"] = device_map
 
             cur_device = get_device_type()
@@ -499,7 +516,7 @@ class ModelLoader:
                 "bnb_4bit_quant_storage": torch.bfloat16,
             }
             if self.cfg.model_config_type in ["jamba", "qwen2_moe"] and not (
-                self.cfg.deepspeed or self.cfg.fsdp
+                self.cfg.deepspeed or self.is_fsdp_enabled
             ):
                 # for some reason, this causes the loss to be off by an order of magnitude
                 # but deepspeed needs this still in bfloat16
@@ -604,9 +621,21 @@ class ModelLoader:
     def _build_model(self) -> bool:
         """Load model, with load strategy depending on config."""
         skip_move_to_device = False
+        if self.is_fsdp_enabled:
+            if self.cfg.fsdp_config.cpu_ram_efficient_loading:
+                skip_move_to_device = True
+                # Don't delete device_map for QLoRA + FSDP - it was set correctly in _set_device_map
+                if (
+                    "device_map" in self.model_kwargs
+                    and not self.is_qlora_and_fsdp_enabled
+                ):
+                    del self.model_kwargs["device_map"]
+            elif self.is_qlora_and_fsdp_enabled:
+                skip_move_to_device = True
+
         if (
-            self.qlora_fsdp
-            and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+            self.is_qlora_and_fsdp_enabled
+            and self.cfg.fsdp_config.cpu_ram_efficient_loading
             and (
                 self.cfg.model_config_type == "dbrx"
                 or self.cfg.qlora_sharded_model_loading
@@ -632,12 +661,6 @@ class ModelLoader:
             and not self.cfg.trust_remote_code
             and not self.cfg.gptq
         ):
-            # TODO: Do we need to open this up for all models?
-            if self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading:
-                skip_move_to_device = True
-                if "device_map" in self.model_kwargs:
-                    del self.model_kwargs["device_map"]
-
             # Please don't remove underscore binding without reading the fn docstring.
             _ = self._configure_zero3_memory_efficient_loading()
 
@@ -691,33 +714,22 @@ class ModelLoader:
                     trust_remote_code=self.cfg.trust_remote_code or False,
                     **self.model_kwargs,
                 )
+        elif self.cfg.gptq:
+            self.model = self.auto_model_loader.from_pretrained(
+                self.base_model,
+                config=self.model_config,
+                trust_remote_code=self.cfg.trust_remote_code or False,
+                **self.model_kwargs,
+            )
         else:
-            if self.cfg.gptq:
-                self.model = self.auto_model_loader.from_pretrained(
-                    self.base_model,
-                    config=self.model_config,
-                    trust_remote_code=self.cfg.trust_remote_code or False,
-                    **self.model_kwargs,
-                )
-            else:
-                if (
-                    self.cfg.fsdp
-                    and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
-                ):
-                    # disabling either of these two still leads to VRAM spike before setting back down
-                    skip_move_to_device = True
-                    if "device_map" in self.model_kwargs:
-                        del self.model_kwargs["device_map"]
-
-                # Please don't remove underscore binding without reading the fn docstring.
-                _ = self._configure_zero3_memory_efficient_loading()
-
-                self.model = self.auto_model_loader.from_pretrained(
-                    self.base_model,
-                    config=self.model_config,
-                    trust_remote_code=self.cfg.trust_remote_code or False,
-                    **self.model_kwargs,
-                )
+            # Please don't remove underscore binding without reading the fn docstring.
+            _ = self._configure_zero3_memory_efficient_loading()
+            self.model = self.auto_model_loader.from_pretrained(
+                self.base_model,
+                config=self.model_config,
+                trust_remote_code=self.cfg.trust_remote_code or False,
+                **self.model_kwargs,
+            )
         if is_deepspeed_zero3_enabled():
             skip_move_to_device = True
 
@@ -753,8 +765,8 @@ class ModelLoader:
             skip_prepare_model_for_kbit_training = True
 
         if (
-            self.qlora_fsdp
-            or (self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading)
+            self.is_qlora_and_fsdp_enabled
+            or (self.is_fsdp_enabled and self.cfg.fsdp_config.cpu_ram_efficient_loading)
             or is_deepspeed_zero3_enabled()
         ):
             # Make sure everything is in the same dtype

src/axolotl/loaders/patch_manager.py

@@ -93,10 +93,14 @@ class PatchManager:
 
     def _apply_fsdp_patches(self):
         """Apply patches for FSDP configurations."""
-        if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
+        if self.cfg.fsdp_config and str(self.cfg.fsdp_version) == "2":
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp2
 
             patch_accelerate_fsdp2()
+            if self.cfg.rl:
+                from axolotl.monkeypatch.trainer.trl import patch_trl_prepare_fsdp2
+
+                patch_trl_prepare_fsdp2()
 
         # if self.cfg.fsdp_config:
         #     # see transformers#39152
@@ -163,13 +167,19 @@ class PatchManager:
 
     def _apply_gradient_checkpointing_patches(self):
         """Apply patches for gradient checkpointing."""
-        if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
+        if (
+            self.cfg.gradient_checkpointing
+            and self.cfg.activation_offloading == "legacy"
+        ):
             from axolotl.monkeypatch.gradient_checkpointing import (
                 hf_grad_checkpoint_offload_wrapper,
             )
 
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
-        if self.cfg.gradient_checkpointing == "offload_disk":
+        elif (
+            self.cfg.gradient_checkpointing
+            and self.cfg.activation_offloading == "offload_disk"
+        ):
             from axolotl.monkeypatch.gradient_checkpointing import (
                 hf_grad_checkpoint_disk_offload_wrapper,
             )

src/axolotl/loaders/utils.py

@@ -195,9 +195,11 @@ def ensure_dtype(model: PreTrainedModel, dtype: torch.dtype = torch.bfloat16):
             bias_mismatch = module.bias.dtype != dtype
 
         if weight_mismatch:
-            print(f"Converting module {name}.weight: {module.weight.dtype} -> {dtype}")
+            LOG.debug(
+                f"Converting module {name}.weight: {module.weight.dtype} -> {dtype}"
+            )
         if bias_mismatch:
-            print(f"Converting module {name}.bias: {module.bias.dtype} -> {dtype}")
+            LOG.debug(f"Converting module {name}.bias: {module.bias.dtype} -> {dtype}")
         if weight_mismatch or bias_mismatch:
             module.to(dtype)
 

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -2,102 +2,65 @@
 monkeypatch for accelerate fsdp2 fix when modifying ordereddict during interation, and saving full state dicts
 """
 
+import copy
+import functools
 import sys
 
 import torch
+from torch import nn
 
+from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 
 
-def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dict):
+def fsdp2_load_full_state_dict(
+    _accelerator, model: torch.nn.Module, full_sd: dict, offload_to_cpu: bool = False
+):
     """
     Loads the full state dict (could be only on rank 0) into the sharded model. This is done by broadcasting the
     parameters from rank 0 to all other ranks. This function modifies the model in-place.
-
     Args:
         accelerator (`Accelerator`): The accelerator instance
         model (`torch.nn.Module`):
             The model to load the state dict into, expected to be on meta device or a VRAM spike can occur
         full_sd (`dict`): The full state dict to load, can only be on rank 0
     """
-    import torch.distributed as dist
     from torch.distributed.tensor import distribute_tensor
 
-    # Model was previously copied to meta device
+    LOG.info("Broadcasting full state dict to all ranks...")
+    import time
+
+    start_time = time.time()
+
     meta_sharded_sd = model.state_dict()
     sharded_sd = {}
-
-    # Rank 0 distributes the full state dict to other ranks
-    def _infer_parameter_dtype(model, param_name, empty_param):
-        try:
-            old_param = model.get_parameter_or_buffer(param_name)
-        except AttributeError:
-            # Need this for LORA, as there some params are not *parameters* of sorts
-            base_param_name, local_param_name = param_name.rsplit(".", 1)
-            submodule = model.get_submodule(base_param_name)
-            old_param = getattr(submodule, local_param_name)
-
-        is_torch_e4m3fn_available = hasattr(torch, "float8_e4m3fn")
-        casting_dtype = None
-        is_param_float8_e4m3fn = (
-            is_torch_e4m3fn_available and empty_param.dtype == torch.float8_e4m3fn
-        )
-
-        if empty_param.dtype.is_floating_point and not is_param_float8_e4m3fn:
-            casting_dtype = old_param.dtype
-
-        return old_param is not None and old_param.is_contiguous(), casting_dtype
-
-    def _cast_and_contiguous(tensor, to_contiguous, dtype):
-        if dtype is not None:
-            tensor = tensor.to(dtype=dtype)
-        if to_contiguous:
-            tensor = tensor.contiguous()
-        return tensor
-
-    param_names = sorted(meta_sharded_sd.keys())
-
-    for param_name in param_names:
-        mesh = meta_sharded_sd[param_name].device_mesh
-        if accelerator.is_main_process:
-            full_param = full_sd[param_name].detach().cuda()
-            dist.broadcast(full_param, src=0, group=mesh.get_group())
-            sharded_tensor = distribute_tensor(
-                full_param, mesh, sharded_sd[param_name].placements
-            )
-            to_contiguous, casting_dtype = _infer_parameter_dtype(
-                model,
-                param_name,
-                full_param,
-            )
-            sharded_tensor = _cast_and_contiguous(
-                sharded_tensor, to_contiguous, casting_dtype
-            )
-            sharded_sd[param_name] = sharded_tensor
-        else:
-            full_tensor = torch.empty(
-                sharded_sd[param_name].size(),
-                device="cuda",
-                dtype=sharded_sd[param_name].dtype,
-            )
-            dist.broadcast(full_tensor, src=0, group=mesh.get_group())
-            sharded_tensor = distribute_tensor(
-                full_tensor, mesh, sharded_sd[param_name].placements
-            )
-            to_contiguous, casting_dtype = _infer_parameter_dtype(
-                model,
-                param_name,
+    for param_name, full_tensor in full_sd.items():
+        sharded_meta_param = meta_sharded_sd.get(param_name)
+        full_tensor = full_tensor.to(sharded_meta_param.dtype).to(torch.device("cuda"))
+        if hasattr(sharded_meta_param, "device_mesh"):
+            sharded_param = distribute_tensor(
                 full_tensor,
+                sharded_meta_param.device_mesh,
+                sharded_meta_param.placements,
+                src_data_rank=0,
             )
-            sharded_tensor = _cast_and_contiguous(
-                sharded_tensor, to_contiguous, casting_dtype
-            )
-            sharded_sd[param_name] = sharded_tensor
+        else:
+            sharded_param = full_tensor
 
-    # we set `assign=True` because our params are on meta device
-    model.load_state_dict(sharded_sd, assign=True)
+        if offload_to_cpu:
+            sharded_param = sharded_param.cpu()
+
+        sharded_sd[param_name] = nn.Parameter(sharded_param)
+        del full_tensor
+        full_sd[param_name] = None
+    model.load_state_dict(sharded_sd, assign=True, strict=True)
+    end_time = time.time()
+    LOG.debug(
+        f"Time taken to load full state dict: {(end_time - start_time):.2f} seconds"
+    )
+    log_gpu_memory_usage(LOG, "Memory usage after broadcasting full state dict", 0)
     return model
 
 
@@ -191,17 +154,195 @@ def get_state_dict(self, model, unwrap=True):
     return state_dict
 
 
-def patch_accelerate_fsdp2():
-    import accelerate
-    from accelerate.utils import fsdp_utils
+def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
+    """Helper function to process LoRA modules for FSDP2."""
+    from torch.distributed.fsdp import fully_shard
 
-    fsdp_utils.fsdp2_load_full_state_dict = fsdp2_load_full_state_dict
-    setattr(
-        sys.modules["accelerate.utils.fsdp_utils"],
-        "fsdp2_load_full_state_dict",
-        fsdp2_load_full_state_dict,
+    log_bias_dtype_mismatch = False
+
+    # Linear4Bit will keep it's bias term in fp32. If the weight dtype is in bf16 we are not able to
+    # wrap this. Therefore we must ensure the bias has the same dtype as the weight
+    if module.base_layer.bias is not None:
+        if module.base_layer.weight.dtype != module.base_layer.bias.dtype:
+            log_bias_dtype_mismatch = True
+            module.base_layer.bias.data = module.base_layer.bias.data.to(
+                module.base_layer.weight.dtype
+            )
+
+    for active_adapter in module.active_adapters:
+        if module.lora_A:
+            fully_shard(module.lora_A[active_adapter], **fsdp2_kwargs)
+        if module.lora_B:
+            fully_shard(module.lora_B[active_adapter], **fsdp2_kwargs)
+        if module.lora_embedding_A:
+            fully_shard(module.lora_embedding_A[active_adapter], **fsdp2_kwargs)
+        if module.lora_embedding_B:
+            fully_shard(module.lora_embedding_B[active_adapter], **fsdp2_kwargs)
+        if module.lora_magnitude_vector:
+            fully_shard(module.lora_magnitude_vector[active_adapter], **fsdp2_kwargs)
+    return log_bias_dtype_mismatch
+
+
+def fsdp2_prepare_model(accelerator, model: torch.nn.Module) -> torch.nn.Module:
+    """Prepares the model for FSDP2 in-place. Also returns the model to avoid misuse of the original model.
+
+    Args:
+        accelerator (`Accelerator`): The accelerator instance
+        model (`torch.nn.Module`): The model to prepare
+
+    Returns:
+        `torch.nn.Module`: Prepared model
+    """
+    from accelerate.utils import get_module_children_bottom_up, is_compiled_module
+    from accelerate.utils.fsdp_utils import fsdp2_prepare_auto_wrap_policy
+    from accelerate.utils.modeling import get_non_persistent_buffers
+    from peft import PeftModel
+    from peft.tuners.lora import LoraLayer
+    from torch.distributed.fsdp import (
+        CPUOffloadPolicy,
+        FSDPModule,
+        MixedPrecisionPolicy,
+        fully_shard,
     )
 
+    is_type_fsdp = isinstance(model, FSDPModule) or (
+        is_compiled_module(model)
+        and isinstance(model._orig_mod, FSDPModule)  # pylint: disable=protected-access
+    )
+    if is_type_fsdp:
+        return model
+
+    fsdp2_plugin = accelerator.state.fsdp_plugin
+
+    original_sd = model.state_dict()
+
+    from torch.distributed.fsdp.wrap import (
+        size_based_auto_wrap_policy,
+        transformer_auto_wrap_policy,
+    )
+
+    # We need the `auto_wrap_policy` original type to create a custom poilicy function for sharding
+    # This is because `fully_shard` doesn't support old auto wrap policies, rather we have to imitate the behaviour
+    if fsdp2_plugin.auto_wrap_policy is transformer_auto_wrap_policy:
+        pass  # auto_wrap_policy_type = "transformer"
+    elif fsdp2_plugin.auto_wrap_policy is size_based_auto_wrap_policy:
+        pass  # auto_wrap_policy_type = "size"
+
+    # We set `auto_wrap_policy` to `functools.partial` to avoid creating it again
+    # This is because of `apply_activation_checkpointing` which will can reuse this function
+    fsdp2_plugin.set_auto_wrap_policy(model)
+
+    if fsdp2_plugin.activation_checkpointing:
+        from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
+            CheckpointImpl,
+            apply_activation_checkpointing,
+            checkpoint_wrapper,
+        )
+
+        # Apply activation checkpointing before applying `fully_shard`
+        apply_activation_checkpointing(
+            model,
+            checkpoint_wrapper_fn=functools.partial(
+                checkpoint_wrapper,
+                checkpoint_impl=CheckpointImpl.NO_REENTRANT,
+            ),
+            auto_wrap_policy=fsdp2_plugin.auto_wrap_policy,
+        )
+
+    fsdp2_kwargs = {
+        "reshard_after_forward": fsdp2_plugin.reshard_after_forward,
+        "offload_policy": fsdp2_plugin.cpu_offload,
+        # `fully_shard` doesn't accept `None` in case of `MixedPrecisionPolicy`
+        "mp_policy": fsdp2_plugin.mixed_precision_policy or MixedPrecisionPolicy(),
+    }
+
+    model_has_params4bit = False
+    for _, param in model.named_parameters():
+        # this is a temporary fix whereby loading models with bnb params cannot be moved from
+        # GPU to a meta device due with FSDP2 because torch operations don't return the original class type
+        # bypassing the move to meta will still cause the VRAM spike, but at least it still will load
+        if param.__class__.__name__ == "Params4bit":
+            model_has_params4bit = True
+            break
+
+    if fsdp2_plugin.cpu_ram_efficient_loading and not model_has_params4bit:
+        # Context: `fully_shard` moves the model to GPU if it was on CPU, however it can also be on `meta` and then it stays there even after `fully_shard`
+        # For this reason, we need to move the model to `meta` device, as then sharding happens on `meta` device
+        # If we kept the model on CPU (`cpu_ram_efficient_loading` has model be on CPU on all ranks, though non-main ranks only have `torch.emtpy`), `fully_shard` would move it to GPU
+        # Afterwards, when we call `fsdp2_load_full_state_dict`, us creating the state_dict would result into briefly having two copies of model state_dict on the GPU -> VRAM spike
+
+        # We need to keep the original non-persistent buffers, as those MAY not be in the state_dict, resulting in them staying on meta device
+        # Also, these buffers aren't getting sharded by default
+        # We get the FQNs of all non-persistent buffers, to re-register them after
+        non_persistent_buffer_fqns = get_non_persistent_buffers(
+            model, recurse=True, fqns=True
+        )
+        original_non_persistent_buffers = copy.deepcopy(
+            {k: v for k, v in model.named_buffers() if k in non_persistent_buffer_fqns}
+        )
+        # We move the model to meta device, as then sharding happens on meta device
+        model = model.to(torch.device("meta"))
+        # We need to re-tie the weights, not exactly sure why, but if we don't do this, reference to `lm_head/embed_tokens` stay hanging -> more VRAM usage
+        # We assume `transformers` models have a `tie_weights` method if they support it
+        if hasattr(model, "tie_weights"):
+            model.tie_weights()
+
+    is_peft_model = isinstance(model, PeftModel)
+
+    auto_wrap_policy = fsdp2_prepare_auto_wrap_policy(fsdp2_plugin, model)
+    log_bias_dtype_mismatch = False
+    if auto_wrap_policy is not None:
+        for module in get_module_children_bottom_up(model)[:-1]:
+            if is_peft_model and isinstance(module, LoraLayer):
+                module_log_bias_mismatch = _process_lora_module_for_fsdp(
+                    module, fsdp2_kwargs
+                )
+                log_bias_dtype_mismatch |= module_log_bias_mismatch
+            if auto_wrap_policy(module) and not isinstance(module, FSDPModule):
+                fully_shard(module, **fsdp2_kwargs)
+
+    fully_shard(model, **fsdp2_kwargs)
+
+    if log_bias_dtype_mismatch:
+        LOG.warning(
+            "Bias dtype mismatch detected in LoRA base linear layer. Bias parameters have been cast to weight dtype."
+        )
+
+    if fsdp2_plugin.cpu_ram_efficient_loading:
+        offload_to_cpu = isinstance(fsdp2_plugin.cpu_offload, CPUOffloadPolicy)
+        fsdp2_load_full_state_dict(
+            accelerator, model, original_sd, offload_to_cpu=offload_to_cpu
+        )
+
+    if fsdp2_plugin.cpu_ram_efficient_loading and not model_has_params4bit:
+        # We re-register the buffers, as they may not be in the state_dict
+        for fqn, buffer_tensor in original_non_persistent_buffers.items():
+            buffer_tensor = buffer_tensor.to(accelerator.device)
+
+            if "." in fqn:
+                parent_fqn, local_buffer_name = fqn.rsplit(".", 1)
+                parent_module = model.get_submodule(parent_fqn)
+            else:
+                local_buffer_name = fqn
+                parent_module = model
+
+            parent_module.register_buffer(
+                local_buffer_name, buffer_tensor, persistent=False
+            )
+
+        # We need to tie the weights again, as call to `load_full_state_dict` breaks the tie
+        # Needs to be called both here and above
+        # removing this call makes the have slightly different loss
+        # removing the call above leads to extra memory usage as explained in the comment above
+        if hasattr(model, "tie_weights"):
+            model.tie_weights()
+    return model
+
+
+def patch_accelerate_fsdp2():
+    import accelerate
+
+    accelerate.accelerator.fsdp2_prepare_model = fsdp2_prepare_model
     accelerate.Accelerator.get_state_dict = get_state_dict
     setattr(
         sys.modules["accelerate"],

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -6,6 +6,10 @@ from typing import Optional, Tuple, Union
 import torch
 import transformers
 
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
 
 def patch_flex_wrapper(**flex_attn_compile_kwargs):
     # TODO remove this patch when transformers#37285 is merged and in a release
@@ -46,10 +50,15 @@ def patch_flex_wrapper(**flex_attn_compile_kwargs):
                 # cause errors. The suggested fix is to compile with "max-autotune-no-cudagraphs"
                 # see https://github.com/pytorch/pytorch/issues/146260 for training
                 self.training = training
+                LOG.info(
+                    "Compiling flex attention with kwargs: %s. This may take a while...",
+                    flex_attn_compile_kwargs,
+                )
                 self._compiled_flex_attention = torch.compile(
                     flex_attention,
                     **flex_attn_compile_kwargs,
                 )
+                LOG.info("Flex attention compiled successfully.")
                 self._is_flex_compiled = True
 
         def __call__(self):

src/axolotl/monkeypatch/gradient_checkpointing/__init__.py

@@ -5,7 +5,7 @@ from functools import partial
 
 from packaging import version
 
-from axolotl.monkeypatch.gradient_checkpointing.offload_cpu import (
+from axolotl.monkeypatch.gradient_checkpointing.offload_cpu import (  # noqa: F401
     CPU_Offloaded_Gradient_Checkpointer,
 )
 from axolotl.monkeypatch.gradient_checkpointing.offload_disk import (

src/axolotl/monkeypatch/gradient_checkpointing/offload_cpu.py

@@ -13,8 +13,17 @@
 # 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.
+
+import inspect
+
 import torch
 from packaging import version
+from torch.utils.checkpoint import (
+    set_device_states,
+)
+
+# support different pytorch versions
+has_device_type = "device_type" in inspect.signature(set_device_states).parameters
 
 torch_version = version.parse(torch.__version__)
 

src/axolotl/monkeypatch/tiled_mlp.py

@@ -1,6 +1,7 @@
 """Monkeypatch for Tiled MLP implementation"""
 
 import math
+import os
 
 import torch
 import torch.distributed as dist
@@ -29,15 +30,18 @@ def patch_tiled_mlp(model_type, use_original_mlp=False, cfg_num_shards=None):
 
             mlp_forward = torch.compile(generic_mlp_forward)
 
+        is_distributed = int(os.environ.get("WORLD_SIZE", 1)) > 1
+
         def tiled_mlp_forward(self, x):
             input_shape = x.shape
             seqlen = input_shape[-2]
             hidden = input_shape[-1]
             if cfg_num_shards is None:
                 num_shards = math.ceil(seqlen / hidden)
-                num_shards_tensor = torch.tensor(num_shards, device=x.device)
-                dist.all_reduce(num_shards_tensor, op=dist.ReduceOp.MAX)
-                num_shards = num_shards_tensor.item()
+                if is_distributed:
+                    num_shards_tensor = torch.tensor(num_shards, device=x.device)
+                    dist.all_reduce(num_shards_tensor, op=dist.ReduceOp.MAX)
+                    num_shards = num_shards_tensor.item()
             else:
                 num_shards = cfg_num_shards
 

src/axolotl/monkeypatch/trainer/trl.py

@@ -0,0 +1,13 @@
+"""Monkeypatch for TRL trainer FSDP preparation."""
+
+
+def prepare_fsdp(model, accelerator):
+    from axolotl.monkeypatch.accelerate.fsdp2 import fsdp2_prepare_model
+
+    return fsdp2_prepare_model(accelerator, model)
+
+
+def patch_trl_prepare_fsdp2():
+    import trl.models.utils
+
+    trl.models.utils.prepare_fsdp = prepare_fsdp

src/axolotl/prompt_strategies/chat_template.py

@@ -379,6 +379,22 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
         Public method that can handle either a single prompt or a batch of prompts.
         """
 
+        def _remove_none_values(obj):
+            """
+            Remove null from a dictionary-like obj or list.
+            These can appear due to Dataset loading causing schema merge.
+            See https://github.com/axolotl-ai-cloud/axolotl/pull/2909
+            """
+            if hasattr(obj, "items"):
+                return {
+                    k: _remove_none_values(v) for k, v in obj.items() if v is not None
+                }
+            if isinstance(obj, list):
+                return [_remove_none_values(elem) for elem in obj]
+            return obj
+
+        prompt = _remove_none_values(prompt)
+
         if not self.is_prompt_batched(prompt) or not self.supports_batched:
             return self._tokenize_single_prompt(prompt)
 

src/axolotl/prompt_strategies/dpo/user_defined.py

@@ -33,7 +33,7 @@ def default(cfg, dataset_idx=0, **kwargs):  # pylint: disable=unused-argument
                 system=sample[field_system], prompt=sample[field_prompt]
             )
         else:
-            sample["prompt"] = prompt_format.format(prompt=sample["prompt"])
+            sample["prompt"] = prompt_format.format(prompt=sample[field_prompt])
         sample["chosen"] = chosen_format.format(chosen=sample[field_chosen])
         sample["rejected"] = rejected_format.format(rejected=sample[field_rejected])
         return sample

src/axolotl/train.py

@@ -15,7 +15,6 @@ from typing import Any, Dict
 
 import torch
 import transformers.modelcard
-from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
 from peft import PeftConfig, PeftModel
@@ -68,7 +67,7 @@ def setup_model_and_tokenizer(
             `None`), and processor (if multimodal, else `None`).
     """
     # Load tokenizer
-    LOG.debug(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    LOG.debug(f"Loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
     tokenizer = load_tokenizer(cfg)
 
     # Load processor for multimodal models if needed
@@ -76,11 +75,8 @@ def setup_model_and_tokenizer(
     if cfg.is_multimodal:
         processor = load_processor(cfg, tokenizer)
 
-    # Load the model and peft_config
-    msg = "loading model"
-    if cfg.adapter:
-        msg += " and peft_config..."
-    LOG.debug(msg)
+    # Load the model
+    LOG.debug("Loading model")
 
     model_loader = ModelLoader(cfg, tokenizer, processor=processor)
     model, peft_config = model_loader.load()
@@ -228,6 +224,9 @@ def execute_training(
         #     torch.set_default_dtype(torch.bfloat16)
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
+        plugin_manager = PluginManager.get_instance()
+        plugin_manager.post_train(cfg, trainer.model)
+
 
 def save_trained_model(
     cfg: DictDefault,
@@ -264,15 +263,6 @@ def save_trained_model(
             "QAT modules have been converted for PTQ. Please ensure you quantize "
             "your model weights with `axolotl quantize`."
         )
-
-    # Handle FSDP state dict type
-    state_dict_type = "FULL_STATE_DICT"
-    if trainer.is_fsdp_enabled and str(cfg.fsdp_config.fsdp_version) != "2":
-        if cfg.fsdp_final_state_dict_type:
-            state_dict_type = cfg.fsdp_final_state_dict_type
-        trainer.accelerator.state.fsdp_plugin.set_state_dict_type(state_dict_type)
-        LOG.info(f"Set FSDP state dict type to {state_dict_type} for saving.")
-
     # Handle ReLoRA early return case
     if cfg.relora_steps:
         if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
@@ -281,22 +271,19 @@ def save_trained_model(
             # final model weights have already been saved by `ReLoRACallback.on_train_end`
             return
 
-    if cfg.fsdp:
-        # TODO: do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
-        # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple
-        # processes attempt to write the same file
-        if (
-            state_dict_type == "SHARDED_STATE_DICT"
-            and cfg.fsdp_config.fsdp_state_dict_type == "SHARDED_STATE_DICT"
-        ):
-            save_fsdp_model(
-                trainer.accelerator.state.fsdp_plugin,
-                trainer.accelerator,
-                trainer.model,
-                cfg.output_dir,
+    if trainer.is_fsdp_enabled:
+        if cfg.fsdp_config or cfg.fsdp:
+            if cfg.fsdp_config.final_state_dict_type:
+                state_dict_type = cfg.fsdp_config.final_state_dict_type
+            else:
+                state_dict_type = cfg.fsdp_config.state_dict_type
+            trainer.accelerator.state.fsdp_plugin.set_state_dict_type(state_dict_type)
+        trainer.save_model(cfg.output_dir)
+        if state_dict_type == "SHARDED_STATE_DICT":
+            LOG.info(
+                "The final model was saved with a sharded state dict. Please ensure you merge "
+                "the sharded weights with `merge-sharded-fsdp-weights`."
             )
-        elif state_dict_type == "FULL_STATE_DICT":
-            trainer.save_model(cfg.output_dir)
     elif cfg.deepspeed and is_deepspeed_zero3_enabled():
         # Copied over from: https://github.com/huggingface/accelerate/blob/5ae611118057232f441055f7ef9ba0b0f2b8d533/docs/source/usage_guides/deepspeed.md#saving-and-loading
         trainer.accelerator.wait_for_everyone()
@@ -526,6 +513,9 @@ def setup_model_and_trainer(cfg: DictDefault, dataset_meta: TrainDatasetMeta) ->
         peft_config=peft_config,
     )
 
+    plugin_manager = PluginManager.get_instance()
+    plugin_manager.post_trainer_create(cfg, trainer)
+
     return (
         trainer,
         model,
@@ -557,9 +547,6 @@ def train(
         processor,
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
-    plugin_manager = PluginManager.get_instance()
-    plugin_manager.post_trainer_create(cfg, trainer)
-
     # Handle untrained tokens if configured
     safe_serialization = cfg.save_safetensors is True
     train_dataset = dataset_meta.train_dataset
@@ -582,6 +569,4 @@ def train(
     if not cfg.use_ray:
         cleanup_distributed()
 
-    plugin_manager.post_train(cfg, model)
-
     return model, tokenizer, trainer

src/axolotl/utils/bench.py

@@ -1,6 +1,7 @@
 """Benchmarking and measurement utilities"""
 
 import functools
+import logging
 
 import torch
 from transformers.utils.import_utils import is_torch_npu_available
@@ -91,21 +92,27 @@ def gpu_memory_usage_smi(device=0):
         return 0.0
 
 
-def log_gpu_memory_usage(log, msg, device):
-    cur_device = get_device_type()
+def log_gpu_memory_usage(
+    log: logging.Logger | logging.LoggerAdapter,
+    msg: str = "",
+    device: int | torch.device = 0,
+):
+    cur_device_type = str(get_device_type())
     if torch.backends.mps.is_available():
         usage, cache, misc = mps_memory_usage_all()
-    elif "npu" in str(cur_device) and is_torch_npu_available():
+    elif "npu" in cur_device_type and is_torch_npu_available():
         usage, cache, misc = npu_memory_usage_all(device)
-    else:
+    elif "gpu" in cur_device_type and torch.cuda.is_available():
         usage, cache, misc = gpu_memory_usage_all(device)
+    else:
+        return
     extras = []
     if cache > 0:
         extras.append(f"+{cache:.03f}GB cache")
     if misc > 0:
         extras.append(f"+{misc:.03f}GB misc")
+    msg = f"{cur_device_type} memory usage:" if not msg else msg
     log.info(
-        f"{str(cur_device)} memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})",
+        f"{msg} {usage:.03f}GB ({', '.join(extras)})",
         stacklevel=2,
     )
-    return usage, cache, misc

src/axolotl/utils/callbacks/__init__.py

@@ -841,21 +841,35 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
 class GCCallback(TrainerCallback):
     """Callback to garbage collect torch cache"""
 
-    def __init__(self, gc_steps=None):
-        self.gc_steps = gc_steps
+    def __init__(self, gc_steps: int | None = -1):
+        self.gc_steps: int = gc_steps or -1
+        self.next_gc_on_begin_step: int = -1
+
+    def _gc(self):
+        torch.cuda.empty_cache()
+        gc.collect()
+
+    def on_step_begin(
+        self, args, state, control, **kwargs  # pylint: disable=unused-argument
+    ):
+        if self.next_gc_on_begin_step == state.global_step:
+            self._gc()
 
     def on_step_end(
         self, args, state, control, **kwargs  # pylint: disable=unused-argument
     ):
-        if self.gc_steps > 0 and state.global_step % self.gc_steps == 0:
-            torch.cuda.empty_cache()
-            gc.collect()
+        if control.should_evaluate:
+            # automatically GC before evals so the eval memory spike from the CEL doesn't OOM the trainer
+            self._gc()
+            # also GC on the start of the next step after the eval
+            self.next_gc_on_begin_step = state.global_step + 1
+        elif self.gc_steps > 0 and state.global_step % self.gc_steps == 0:
+            self._gc()
 
     def on_epoch_end(
         self, args, state, control, **kwargs  # pylint: disable=unused-argument
     ):
-        torch.cuda.empty_cache()
-        gc.collect()
+        self._gc()
 
 
 def colab_inference_post_train_callback(trainer: Trainer):

src/axolotl/utils/callbacks/profiler.py

@@ -19,9 +19,27 @@ class PytorchProfilerCallback(TrainerCallback):
     PyTorch Profiler callback to create snapshots of GPU memory usage at specified steps.
     """
 
-    def __init__(self, steps_to_profile: int = 5):
-        self.steps_to_profile = steps_to_profile
-        if self.steps_to_profile:
+    def __init__(self, steps_to_profile: int = 5, profiler_steps_start: int = 0):
+        # steps are 0 indexed, so to start at 0-th step, we start at beginning of first step,
+        # and finish at end of last step, so 5 steps_to_profile is steps [0, 1, 2, 3, 4]
+        self.profiler_steps_end = profiler_steps_start + steps_to_profile - 1
+        if profiler_steps_start == 0:
+            # start recording memory allocations before everything is allocated, because if we start
+            # at the beginning of step 0, we won't have any memory allocations in the traces
+            torch.cuda.memory._record_memory_history(  # pylint: disable=protected-access
+                enabled="all"
+            )
+            profiler_steps_start = -1
+        self.profiler_steps_start = profiler_steps_start
+
+    def on_step_begin(  # pylint: disable=unused-argument
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        if state.global_step == self.profiler_steps_start:
             torch.cuda.memory._record_memory_history(  # pylint: disable=protected-access
                 enabled="all"
             )
@@ -33,7 +51,28 @@ class PytorchProfilerCallback(TrainerCallback):
         control: TrainerControl,  # pylint: disable=unused-argument
         **kwargs,  # pylint: disable=unused-argument
     ):
-        if state.global_step == self.steps_to_profile:
+        if state.global_step == self.profiler_steps_end:
+            snapshot = torch.cuda.memory._snapshot()  # pylint: disable=protected-access
+            with open(Path(args.output_dir) / "snapshot.pickle", "wb") as fout:
+                dump(snapshot, fout)
+
+            # tell CUDA to stop recording memory allocations now
+            torch.cuda.memory._record_memory_history(  # pylint: disable=protected-access
+                enabled=None
+            )
+
+    def on_train_end(  # pylint: disable=unused-argument
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        # make sure to record if we happen to have more steps than steps to profile
+        if (
+            state.global_step >= self.profiler_steps_start
+            and state.global_step < self.profiler_steps_end
+        ):
             snapshot = torch.cuda.memory._snapshot()  # pylint: disable=protected-access
             with open(Path(args.output_dir) / "snapshot.pickle", "wb") as fout:
                 dump(snapshot, fout)

src/axolotl/utils/config/__init__.py

@@ -116,9 +116,10 @@ def normalize_config(cfg):
     ]
     choose_device(cfg)
     cfg.ddp = cfg.ddp if cfg.ddp is not None else cfg.world_size != 1
-    if cfg.ddp:
+    if cfg.world_size != 1:
         cfg.device_map = {"": int(os.environ.get("LOCAL_RANK", 0))}
-        cfg.batch_size = cfg.batch_size * cfg.world_size
+        if cfg.fsdp or cfg.fsdp_config or cfg.ddp:
+            cfg.batch_size = cfg.batch_size * cfg.world_size
 
     if not cfg.use_ray:
         # delay resolving dtype until on worker node when launching with ray
@@ -274,7 +275,7 @@ def validate_config(
     # Convert datasets to proper format if needed
     if cfg.get("datasets"):
         for idx, ds_cfg in enumerate(cfg["datasets"]):
-            if cfg.get("rl") == "dpo" and not isinstance(ds_cfg, DPODataset):
+            if cfg.get("rl") in ["dpo", "simpo"] and not isinstance(ds_cfg, DPODataset):
                 cfg["datasets"][idx] = DPODataset(**ds_cfg)
             elif cfg.get("rl") == "kto" and not isinstance(ds_cfg, KTODataset):
                 cfg["datasets"][idx] = KTODataset(**dict(ds_cfg))