remove xformers changes

.github/workflows/base.yml

@@ -29,11 +29,11 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
-          - cuda: "126"
-            cuda_version: 12.6.3
+          - cuda: "124"
+            cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.6.0
@@ -43,7 +43,7 @@ jobs:
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.7.0
+            pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
           - cuda: "126"

.github/workflows/main.yml

@@ -15,15 +15,15 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 124
+            cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.7.0
+            pytorch: 2.6.0
             axolotl_extras: vllm
           - cuda: 126
             cuda_version: 12.6.3
@@ -82,17 +82,17 @@ jobs:
     strategy:
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.5.1
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
             is_latest: true
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"

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

@@ -33,6 +33,13 @@ jobs:
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.5.1
+            axolotl_extras:
+            num_gpus: 2
+            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"

.github/workflows/nightlies.yml

@@ -12,6 +12,11 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.5.1
+            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -63,10 +68,10 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 124
+            cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:

.github/workflows/preview-docs.yml

@@ -28,8 +28,6 @@ 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
@@ -52,11 +50,10 @@ jobs:
 
       - name: Netlify Publish
         uses: nwtgck/actions-netlify@v3.0
-        id: netlify
         with:
           publish-dir: './_site'
-          enable-pull-request-comment: false
-          enable-github-deployment: false
+          enable-pull-request-comment: true
+          enable-github-deployment: true
           github-token: ${{ secrets.GITHUB_TOKEN }}
           deploy-message: "Deployed On Netlify"
           github-deployment-environment: 'preview'
@@ -64,13 +61,3 @@ 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 }}

.github/workflows/tests-nightly.yml

@@ -26,7 +26,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -80,9 +80,9 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v --durations=10 tests/patched/
-          pytest -v --durations=10 tests/cli/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v tests/patched/
+          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |

.github/workflows/tests.yml

@@ -52,7 +52,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
     timeout-minutes: 20
 
     steps:
@@ -102,9 +102,9 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
-          pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
-          pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
@@ -125,7 +125,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
     timeout-minutes: 20
 
     steps:
@@ -175,9 +175,9 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v --durations=10 tests/patched/
-          pytest -v --durations=10 tests/cli/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v tests/patched/
+          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -198,7 +198,7 @@ jobs:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras:
           - cuda: 126
@@ -252,6 +252,18 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
+            axolotl_extras: llmcompressor
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.5.1
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 1
             axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1

.runpod/src/config/config.yaml

@@ -97,7 +97,7 @@
 #       # 'no_input_format' cannot include {input}
 #       no_input_format: "{instruction} "
 
-#       # For `completion` datasets only, uses the provided field instead of `text` column
+#       # For `completion` datsets 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

README.md

@@ -55,7 +55,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.6.0
+- PyTorch ≥2.5.1
 
 ### Installation
 

cicd/multigpu.py

@@ -24,9 +24,9 @@ df_template = template_env.get_template("Dockerfile.jinja")
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
-    "CUDA": os.environ.get("CUDA", "126"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
+    "CUDA": os.environ.get("CUDA", "124"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),

cicd/single_gpu.py

@@ -24,9 +24,9 @@ df_template = template_env.get_template(dockerfile)
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
-    "CUDA": os.environ.get("CUDA", "126"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
+    "CUDA": os.environ.get("CUDA", "124"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),

docs/dataset-formats/conversation.qmd

@@ -187,7 +187,6 @@ 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": "...",
@@ -200,7 +199,6 @@ 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

@@ -34,9 +34,9 @@ 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`
+- `main-base-py3.11-cu124-2.5.1`
 
 ## Main
 
@@ -76,12 +76,12 @@ 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-py3.11-cu124-2.5.1`
 - `main-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
-- `main-20250303-py3.11-cu126-2.6.0`
+- `main-20250303-py3.11-cu124-2.5.1`
 - `0.10.1`
 
 ## Cloud

docs/installation.qmd

@@ -15,7 +15,7 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 - NVIDIA GPU (Ampere architecture or newer for `bf16` and Flash Attention) or AMD GPU
 - Python ≥3.11
-- PyTorch ≥2.6.0
+- PyTorch ≥2.5.1
 
 ## Installation Methods {#sec-installation-methods}
 

docs/multi-gpu.qmd

@@ -23,6 +23,8 @@ 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:
@@ -30,6 +32,7 @@ Add to your YAML config:
 ```{.yaml}
 deepspeed: deepspeed_configs/zero1.json
 ```
+
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
@@ -63,75 +66,9 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-::: {.callout-tip}
+## FSDP {#sec-fsdp}
 
-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.
-
-:::
+### Basic FSDP Configuration {#sec-fsdp-config}
 
 ```{.yaml}
 fsdp:
@@ -143,7 +80,6 @@ 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_version: 2
+fsdp:
+  - full_shard
+  - auto_wrap
 fsdp_config:
-  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
+  fsdp_offload_params: true
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
 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

@@ -17,6 +17,7 @@ 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
@@ -274,14 +275,15 @@ rl: dpo
 datasets:
   - path: ...
     split: train
-    type:
-      field_prompt: "prompt"
-      field_system: "system"
-      field_chosen: "chosen"
-      field_rejected: "rejected"
-      prompt_format: "{prompt}"
-      chosen_format: "{chosen}"
-      rejected_format: "{rejected}"
+    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}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.
@@ -474,13 +476,14 @@ rl: kto
 datasets:
   - path: ...
     split: train
-    type:
-      field_prompt: "prompt"
-      field_system: "system"
-      field_completion: "completion"
-      field_label: "label"
-      prompt_format: "{prompt}"
-      completion_format: "{completion}"
+    type: user_defined.default
+
+    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

@@ -1,5 +0,0 @@
-# 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

@@ -1,70 +0,0 @@
-# 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) 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.
-
-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
-
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Devstral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
-
-    Here is an example of how to install from main for pip:
-
-```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0+)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
-```
-
-2. Run the finetuning example:
-
-```bash
-axolotl train examples/devstral/devstral-small-qlora.yml
-```
-
-This config uses about 21GB VRAM.
-
-Let us know how it goes. Happy finetuning! 🚀
-
-### TIPS
-
-- 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
-
-- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
-- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
-- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
-- [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy)
-- [Liger Kernel](https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels)
-
-## Limitations
-
-We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
-
-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)
-- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)
-
-
-## Future Work
-
-- Add parity to Preference Tuning, RL, Multi-modal, etc.
-- Add parity to other tokenizer configs like overriding tokens.

examples/devstral/devstral-small-qlora.yml

@@ -1,64 +0,0 @@
-base_model: mistralai/Devstral-Small-2507
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-load_in_8bit: false
-load_in_4bit: true
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/qlora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-pad_to_sequence_len: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0
-lora_target_linear: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
-warmup_ratio: 0.05
-evals_per_epoch: 4
-saves_per_epoch: 1
-
-weight_decay: 0.0
-special_tokens:

examples/lfm2/README.md

@@ -1,7 +0,0 @@
-# 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

@@ -1,48 +0,0 @@
-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

examples/magistral/README.md

@@ -18,10 +18,16 @@ git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+pip3 install --no-build-isolation -e '.[flash-attn,mistral]'
 ```
 
-2. Run the finetuning example:
+2. Download the example config:
+
+```bash
+axolotl fetch examples
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/magistral/magistral-small-qlora.yaml
@@ -36,7 +42,7 @@ Let us know how it goes. Happy finetuning! 🚀
 - For inference, the official MistralAI team recommends `top_p: 0.95` and `temperature: 0.7` with `max_tokens: 40960`.
 - 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).
+- The dataset format is the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
 
 ## Optimization Guides
 
@@ -48,7 +54,7 @@ Let us know how it goes. Happy finetuning! 🚀
 
 We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
 
-In addition, we do not support overriding tokens yet.
+The tokenizer does not work with `dataset.map` with multiprocessing, so we had to disable it. In addition, we do not support overriding tokens yet.
 
 ## Related Resources
 

requirements.txt

@@ -11,7 +11,7 @@ liger-kernel==0.5.10
 
 packaging==23.2
 
-huggingface_hub==0.32.2
+huggingface_hub[hf_xet]==0.33.0
 peft==0.15.2
 transformers==4.53.1
 tokenizers>=0.21.1
@@ -68,4 +68,4 @@ schedulefree==1.4.1
 axolotl-contribs-lgpl==0.0.6
 axolotl-contribs-mit==0.0.3
 
-mistral-common==1.7.0
+mistral-common==1.6.3

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@865b899"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@622068a"'
 )

setup.py

@@ -65,18 +65,13 @@ def parse_requirements(extras_require_map):
                 raise ValueError("Invalid version format")
 
             if (major, minor) >= (2, 7):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                if patch == 0:
-                    _install_requires.append("xformers==0.0.30")
-                else:
-                    _install_requires.append("xformers==0.0.31.post1")
-                extras_require_map["vllm"] = ["vllm>=0.9.0"]
+                # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
+                extras_require_map["vllm"] = ["vllm==0.9.2"]
             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
-                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
+                extras_require_map["vllm"] = ["vllm==0.9.2"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:

src/axolotl/__init__.py

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

src/axolotl/cli/config.py

@@ -16,7 +16,6 @@ from transformers.utils import is_torch_bf16_gpu_available
 from axolotl.integrations.base import PluginManager
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
-    migrate_fsdp_config,
     normalize_cfg_datasets,
     normalize_config,
     validate_config,
@@ -227,7 +226,6 @@ def load_cfg(
         },
     )
 
-    migrate_fsdp_config(cfg)
     prepare_optim_env(cfg)
     prepare_opinionated_env(cfg)
     normalize_config(cfg)

src/axolotl/cli/train.py

@@ -109,13 +109,6 @@ 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/core/builders/base.py

@@ -501,10 +501,6 @@ 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)

src/axolotl/core/builders/causal.py

@@ -151,6 +151,14 @@ 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
 

src/axolotl/core/builders/rl.py

@@ -208,7 +208,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             callbacks=self.get_callbacks(),
             **trainer_kwargs,
         )
-        if self.cfg.fsdp_config or self.cfg.fsdp:
+        if 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)
@@ -218,3 +218,21 @@ 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

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

src/axolotl/core/trainers/trl.py

@@ -1,9 +1,12 @@
-"""Module for TRL RL trainers"""
+"""Module for TRL PPO trainer"""
 
+import torch
+from tqdm import tqdm
 from trl import (
     CPOTrainer,
     KTOTrainer,
     ORPOTrainer,
+    PPOTrainer,
     PRMTrainer,
     RewardTrainer,
 )
@@ -13,6 +16,64 @@ 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/datasets.py

@@ -48,6 +48,13 @@ class TokenizedPromptDataset(Dataset):
         features = dataset.features.keys()
         num_proc = min(64, self.process_count if self.process_count else os.cpu_count())
 
+        # Disable multiprocessing if the tokenizer doesn't support it (e.g., mistral_common)
+        if not getattr(self.prompt_tokenizer, "supports_multiprocessing", True):
+            LOG.info(
+                "Disabling multiprocessing for tokenizer as it doesn't support it (e.g., mistral_common)"
+            )
+            num_proc = 1
+
         map_kwargs = {}
         if self.prompt_tokenizer.supports_batched:
             map_kwargs["batched"] = True

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@865b899"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@622068a"
 ```
 
 ## Usage

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -32,7 +32,7 @@ LOG = get_logger(__name__)
 
 _CCE_INSTALL_MESSAGE = (
     "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@865b899"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@622068a"`'
 )
 
 

src/axolotl/integrations/kd/README.md

@@ -11,7 +11,7 @@ kd_ce_alpha: 0.1
 kd_alpha: 0.9
 kd_temperature: 1.0
 
-torch_compile: True  # torch>=2.6.0, recommended to reduce vram
+torch_compile: True  # torch>=2.5.1, recommended to reduce vram
 
 datasets:
   - path: ...

src/axolotl/loaders/adapter.py

@@ -122,9 +122,9 @@ def load_lora(
     rank = int(os.environ.get("LOCAL_RANK", 0))
 
     if (
-        cfg.fsdp_config
+        cfg.fsdp
         and cfg.adapter
-        and cfg.fsdp_config.cpu_ram_efficient_loading
+        and cfg.fsdp_config.fsdp_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_config
+        cfg.fsdp
         and cfg.adapter
-        and cfg.fsdp_config.cpu_ram_efficient_loading
+        and cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
         and rank != 0
     ):
         setup_quantized_peft_meta_for_training(model)

src/axolotl/loaders/model.py

@@ -140,15 +140,10 @@ class ModelLoader:
         """Check if flash attention is installed."""
         return find_spec("flash_attn") is not None
 
-    @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):
+    @cached_property
+    def qlora_fsdp(self):
         """Property that determines if FSDP with QLoRA is enabled."""
-        return self.is_fsdp_enabled and self.cfg.adapter == "qlora"
+        return self.cfg.fsdp and self.cfg.adapter == "qlora"
 
     def load(self) -> tuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]:
         """Load and prepare the model with all configurations and patches.
@@ -194,15 +189,15 @@ class ModelLoader:
         # Handle PeftModel if needed
         if (
             isinstance(self.model, (peft.PeftModel, peft.PeftModelForCausalLM))
-            and not self.is_qlora_and_fsdp_enabled
+            and not self.qlora_fsdp
         ):
             self.model = self.model.merge_and_unload()
 
         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 _resize_token_embeddings(self):
         """Resize token embeddings if needed."""
@@ -256,13 +251,22 @@ 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.is_fsdp_enabled:
+        if not self.cfg.fsdp:
             # 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:
@@ -278,7 +282,7 @@ class ModelLoader:
             self._set_z3_leaf_modules()
 
         # Apply gradient checkpointing if needed
-        needs_fa2_dtype = self.cfg.adapter or self.is_fsdp_enabled
+        needs_fa2_dtype = self.cfg.adapter or self.cfg.fsdp
         if self.cfg.adapter in ["lora", "qlora"]:
             needs_fa2_dtype = True
             if self.cfg.gradient_checkpointing:
@@ -294,12 +298,10 @@ 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.is_qlora_and_fsdp_enabled
-            )
-            or (
-                # CCE requires embedding layers to be in fp16/bf16 for backward pass
-                self.cfg.cut_cross_entropy
+                and not self.qlora_fsdp
             )
+            # CCE requires embedding layers to be in fp16/bf16 for backward pass
+            or self.cfg.cut_cross_entropy
         )
 
         if should_convert:
@@ -355,6 +357,7 @@ 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:
@@ -427,17 +430,7 @@ class ModelLoader:
 
         self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
 
-        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:
+        if not is_deepspeed_zero3_enabled():
             self.model_kwargs["device_map"] = device_map
 
             cur_device = get_device_type()
@@ -506,7 +499,7 @@ 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.is_fsdp_enabled
+                self.cfg.deepspeed or self.cfg.fsdp
             ):
                 # for some reason, this causes the loss to be off by an order of magnitude
                 # but deepspeed needs this still in bfloat16
@@ -611,21 +604,9 @@ 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.is_qlora_and_fsdp_enabled
-            and self.cfg.fsdp_config.cpu_ram_efficient_loading
+            self.qlora_fsdp
+            and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
             and (
                 self.cfg.model_config_type == "dbrx"
                 or self.cfg.qlora_sharded_model_loading
@@ -651,6 +632,12 @@ 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()
 
@@ -704,22 +691,33 @@ 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:
-            # 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 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,
+                )
         if is_deepspeed_zero3_enabled():
             skip_move_to_device = True
 
@@ -755,8 +753,8 @@ class ModelLoader:
             skip_prepare_model_for_kbit_training = True
 
         if (
-            self.is_qlora_and_fsdp_enabled
-            or (self.is_fsdp_enabled and self.cfg.fsdp_config.cpu_ram_efficient_loading)
+            self.qlora_fsdp
+            or (self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading)
             or is_deepspeed_zero3_enabled()
         ):
             # Make sure everything is in the same dtype

src/axolotl/loaders/patch_manager.py

@@ -7,7 +7,6 @@ import importlib.util
 from functools import cached_property
 
 import addict
-import torch
 import transformers
 from transformers import PretrainedConfig, PreTrainedModel
 
@@ -94,14 +93,10 @@ class PatchManager:
 
     def _apply_fsdp_patches(self):
         """Apply patches for FSDP configurations."""
-        if self.cfg.fsdp_config and str(self.cfg.fsdp_version) == "2":
+        if self.cfg.fsdp_config and str(self.cfg.fsdp_config.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
@@ -170,25 +165,10 @@ class PatchManager:
         """Apply patches for gradient checkpointing."""
         if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
             from axolotl.monkeypatch.gradient_checkpointing import (
-                CheckpointFunctionWithCPUOffload,
                 hf_grad_checkpoint_offload_wrapper,
             )
 
-            if (
-                self.cfg.gradient_checkpointing_kwargs
-                and "use_reentrant" in self.cfg.gradient_checkpointing_kwargs
-                and self.cfg.gradient_checkpointing_kwargs["use_reentrant"] is False
-            ):
-                transformers.modeling_utils.checkpoint = (
-                    hf_grad_checkpoint_offload_wrapper
-                )
-            else:
-                transformers.modeling_utils.checkpoint.CheckpointFunction = (
-                    CheckpointFunctionWithCPUOffload
-                )
-                torch.utils.checkpoint.CheckpointFunction = (
-                    CheckpointFunctionWithCPUOffload
-                )
+            transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
         if self.cfg.gradient_checkpointing == "offload_disk":
             from axolotl.monkeypatch.gradient_checkpointing import (
                 hf_grad_checkpoint_disk_offload_wrapper,
@@ -398,18 +378,6 @@ class PatchManager:
                     "Shifted-sparse attention not currently implemented without flash attention."
                 )
 
-        from axolotl.monkeypatch.llama_attn_hijack_flash import (
-            is_xformers_swiglu_available,
-        )
-
-        if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
-            from axolotl.monkeypatch.llama_attn_hijack_flash import (
-                patch_mlp_with_swiglu,
-            )
-
-            LOG.info("Patching with SwiGLU...")
-            patch_mlp_with_swiglu(self.cfg.model_config_type)
-
     def _apply_llama_flash_attn_patches(self, model):
         """Apply LLaMA-specific flash attention patches."""
         if (
@@ -420,14 +388,15 @@ class PatchManager:
             and not self.inference
         ):
             # TODO(MengqingCao): split these patches seperately
-            from axolotl.monkeypatch.llama_attn_hijack_flash import (  # is_xformers_swiglu_available,; replace_llama_mlp_with_swiglu,
+            from axolotl.monkeypatch.llama_attn_hijack_flash import (
+                is_xformers_swiglu_available,
+                replace_llama_mlp_with_swiglu,
                 replace_llama_qkv_with_fused,
             )
 
-            # if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
-            #     LOG.info("Patching with SwiGLU...")
-            #     # replace_llama_mlp_with_swiglu(model)
-            #     patch_mlp_with_swiglu(model)
+            if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
+                LOG.info("Patching with SwiGLU...")
+                replace_llama_mlp_with_swiglu(model)
 
             if self.cfg.flash_attn_fuse_qkv:
                 LOG.info("Patching with fused QKV...")

src/axolotl/loaders/utils.py

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

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -2,65 +2,102 @@
 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, offload_to_cpu: bool = False
-):
+def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dict):
     """
     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
 
-    LOG.info("Broadcasting full state dict to all ranks...")
-    import time
-
-    start_time = time.time()
-
+    # Model was previously copied to meta device
     meta_sharded_sd = model.state_dict()
     sharded_sd = {}
-    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,
+
+    # 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:
-            sharded_param = full_tensor
+            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,
+                full_tensor,
+            )
+            sharded_tensor = _cast_and_contiguous(
+                sharded_tensor, to_contiguous, casting_dtype
+            )
+            sharded_sd[param_name] = sharded_tensor
 
-        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)
+    # we set `assign=True` because our params are on meta device
+    model.load_state_dict(sharded_sd, assign=True)
     return model
 
 
@@ -154,195 +191,17 @@ def get_state_dict(self, model, unwrap=True):
     return state_dict
 
 
-def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
-    """Helper function to process LoRA modules for FSDP2."""
-    from torch.distributed.fsdp import fully_shard
-
-    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
+    from accelerate.utils import fsdp_utils
+
+    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,
+    )
 
-    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

@@ -6,10 +6,6 @@ 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
@@ -50,15 +46,10 @@ 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

@@ -5,8 +5,7 @@ from functools import partial
 
 from packaging import version
 
-from axolotl.monkeypatch.gradient_checkpointing.offload_cpu import (  # noqa: F401
-    CheckpointFunctionWithCPUOffload,
+from axolotl.monkeypatch.gradient_checkpointing.offload_cpu import (
     CPU_Offloaded_Gradient_Checkpointer,
 )
 from axolotl.monkeypatch.gradient_checkpointing.offload_disk import (

src/axolotl/monkeypatch/gradient_checkpointing/offload_cpu.py

@@ -13,24 +13,8 @@
 # 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 contextlib
-import inspect
-
 import torch
 from packaging import version
-from torch.utils.checkpoint import (
-    _get_autocast_kwargs,
-    _get_device_module,
-    _infer_device_type,
-    check_backward_validity,
-    detach_variable,
-    get_device_states,
-    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__)
 
@@ -76,153 +60,3 @@ class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
         ) + (
             None,
         ) * len(ctx.args)
-
-
-# Copyright 2025 Snowflake Inc.
-# SPDX-License-Identifier: Apache-2.0
-# https://github.com/snowflakedb/ArcticTraining/blob/main/arctic_training/monkey_patches.py
-class CheckpointFunctionWithCPUOffload(torch.autograd.Function):
-    """
-    This is a torch/utils/checkpoint.py CheckpointFunction monkey patch that offloads the first tensor to cpu during forward and back to cuda during backward. This allows significant memory savings when using a very long seqlen. e.g. for llama 8b at 100k it's 24GB saved per gpu: `((100_000*4096)*2*32/2**30)`
-    In the case of a very long seqlen 100k+ the copying to/from cpu overhead is not big, because dense quadratic attention compute will dominate.
-    """
-
-    @staticmethod
-    def forward(ctx, run_function, preserve_rng_state, *args):
-        check_backward_validity(args)
-        ctx.run_function = run_function
-        ctx.preserve_rng_state = preserve_rng_state
-        # Accommodates the (remote) possibility that autocast is enabled for cpu AND gpu.
-        ctx.device_type = _infer_device_type(*args)
-        ctx.device_autocast_kwargs, ctx.cpu_autocast_kwargs = _get_autocast_kwargs(
-            ctx.device_type
-        )
-        if preserve_rng_state:
-            ctx.fwd_cpu_state = torch.get_rng_state()
-            # Don't eagerly initialize the cuda context by accident.
-            # (If the user intends that the context is initialized later, within their
-            # run_function, we SHOULD actually stash the cuda state here.  Unfortunately,
-            # we have no way to anticipate this will happen before we run the function.)
-            ctx.had_device_in_fwd = False
-            device_module = _get_device_module(ctx.device_type)
-            if getattr(device_module, "_initialized", False):
-                ctx.had_device_in_fwd = True
-                ctx.fwd_devices, ctx.fwd_device_states = get_device_states(*args)
-
-        # Save non-tensor inputs in ctx, keep a placeholder None for tensors
-        # to be filled out during the backward.
-        ctx.inputs = []
-        ctx.tensor_indices = []
-        tensor_inputs = []
-        # x = None
-        for i, arg in enumerate(args):
-            if torch.is_tensor(arg):
-                # cpu-offload
-                # we don't want the 2nd tensor - usually it's a shared 4D attn mask which is huge [seq,seq]
-                # upstream could accept a list of arg indices to offload
-                if i == 0:
-                    # print(f"{arg.shape=}")
-                    ctx.x_device = arg.device
-                    ctx.x_requires_grad = arg.requires_grad
-                    t = arg.detach().cpu()
-                else:
-                    t = arg
-                tensor_inputs.append(t)
-                ctx.tensor_indices.append(i)
-                ctx.inputs.append(None)
-            else:
-                ctx.inputs.append(arg)
-
-        ctx.save_for_backward(*tensor_inputs)
-
-        with torch.no_grad():
-            outputs = run_function(*args)
-
-        return outputs
-
-    @staticmethod
-    def backward(ctx, *args):
-        if (
-            not torch.autograd._is_checkpoint_valid()  # pylint: disable=protected-access
-        ):
-            raise RuntimeError(
-                "When use_reentrant=True, torch.utils.checkpoint is incompatible"
-                " with .grad() or passing an `inputs` parameter to .backward()."
-                " To resolve this error, you can either set use_reentrant=False,"
-                " or call .backward() without passing the `inputs` argument."
-            )
-        # Copy the list to avoid modifying original list.
-        inputs = list(ctx.inputs)
-        tensor_indices = ctx.tensor_indices
-        tensors = ctx.saved_tensors
-
-        # Fill in inputs with appropriate saved tensors.
-        for i, idx in enumerate(tensor_indices):
-            if i == 0:
-                t = (
-                    tensors[i]
-                    .to(ctx.x_device)
-                    .detach()
-                    .requires_grad_(ctx.x_requires_grad)
-                )
-            else:
-                t = tensors[i]
-            inputs[idx] = t
-
-        # Stash the surrounding rng state, and mimic the state that was
-        # present at this time during forward.  Restore the surrounding state
-        # when we're done.
-        rng_devices = []
-        if ctx.preserve_rng_state and ctx.had_device_in_fwd:
-            rng_devices = ctx.fwd_devices
-        with torch.random.fork_rng(
-            devices=rng_devices,
-            enabled=ctx.preserve_rng_state,
-            device_type=ctx.device_type,
-        ):
-            if ctx.preserve_rng_state:
-                torch.set_rng_state(ctx.fwd_cpu_state)
-                if ctx.had_device_in_fwd:
-                    if has_device_type:
-                        # newer pytorch (as early as 2.7)
-                        set_device_states(
-                            ctx.fwd_devices,
-                            ctx.fwd_device_states,
-                            device_type=ctx.device_type,
-                        )
-                    else:
-                        # older pytorch (at least 2.4)
-                        set_device_states(ctx.fwd_devices, ctx.fwd_device_states)
-            detached_inputs = detach_variable(tuple(inputs))
-
-            device_autocast_ctx = (
-                torch.amp.autocast(
-                    device_type=ctx.device_type, **ctx.device_autocast_kwargs
-                )
-                if torch.amp.is_autocast_available(ctx.device_type)
-                else contextlib.nullcontext()
-            )
-            with torch.enable_grad(), device_autocast_ctx, torch.amp.autocast("cpu", **ctx.cpu_autocast_kwargs):  # type: ignore[attr-defined]
-                outputs = ctx.run_function(*detached_inputs)
-
-        if isinstance(outputs, torch.Tensor):
-            outputs = (outputs,)
-
-        # run backward() with only tensor that requires grad
-        outputs_with_grad = []
-        args_with_grad = []
-        for i in range(len(outputs)):  # pylint: disable=consider-using-enumerate
-            if torch.is_tensor(outputs[i]) and outputs[i].requires_grad:
-                outputs_with_grad.append(outputs[i])
-                args_with_grad.append(args[i])
-        if len(outputs_with_grad) == 0:
-            raise RuntimeError(
-                "none of output has requires_grad=True, this checkpoint() is not necessary"
-            )
-        torch.autograd.backward(outputs_with_grad, args_with_grad)
-        grads = tuple(
-            inp.grad if isinstance(inp, torch.Tensor) else None
-            for inp in detached_inputs
-        )
-
-        return (None, None) + grads

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -82,28 +82,6 @@ def replace_llama_mlp_with_swiglu(model):
             set_module_name(model, name, mlp)
 
 
-def patch_mlp_with_swiglu(model_type):
-    if is_xformers_swiglu_available():
-        from axolotl.monkeypatch.xformers_ import FusedMLPv2 as FusedMLP
-    else:
-        raise RuntimeError("xformers SwiGLU not available for this environment")
-
-    try:
-        # Dynamically import the module and MLP class
-        module_path = f"transformers.models.{model_type}.modeling_{model_type}"
-        model_cls_prefix = "".join(
-            [part.capitalize() for part in model_type.split("_")]
-        )
-        module = __import__(module_path, fromlist=[f"{model_cls_prefix}MLP"])
-        _ = getattr(module, f"{model_cls_prefix}MLP")
-        setattr(module, f"{model_cls_prefix}MLP", FusedMLP)
-    except (ImportError, AttributeError) as e:
-        raise RuntimeError(
-            f"Could not import MLP class for model_type: {model_type}. "
-            f"Error: {str(e)}"
-        ) from e
-
-
 def replace_llama_qkv_with_fused(model):
     for name, module in model.named_modules():
         if isinstance(module, LlamaAttention):

src/axolotl/monkeypatch/multipack.py

@@ -35,7 +35,6 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "deepseek_v3",
     "glm",
     "glm4",
-    "smollm3",
 ]
 
 

src/axolotl/monkeypatch/tiled_mlp.py

@@ -1,7 +1,6 @@
 """Monkeypatch for Tiled MLP implementation"""
 
 import math
-import os
 
 import torch
 import torch.distributed as dist
@@ -30,18 +29,15 @@ 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)
-                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()
+                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

@@ -1,13 +0,0 @@
-"""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/monkeypatch/xformers_/__init__.py

@@ -1,11 +1,8 @@
 """
 Fused MLP layer for incrementally improved training efficiency
 """
-from collections import OrderedDict
 
 import torch
-from torch import nn
-from transformers.activations import ACT2FN
 from transformers.models.llama.modeling_llama import LlamaMLP
 from xformers.ops import SwiGLU
 
@@ -53,129 +50,3 @@ class FusedMLP(torch.nn.Module):
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
         return self.swiglu(x)
-
-class FusedMLPv2(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        self.swiglu = SwiGLU(
-            in_features=self.hidden_size,
-            hidden_features=self.intermediate_size,
-            bias=config.mlp_bias,
-            _pack_weights=True,
-        )
-        assert config.hidden_act == "silu"
-
-    def _convert_unpacked_to_packed_state_dict(self, unpacked_state_dict):
-        """
-        Convert state dict from unpacked format (w1, w2, w3) to packed format (w13, w2).
-        """
-        packed_state_dict = OrderedDict()
-
-        # Handle w1 and w3 -> w13 conversion for weights
-        if 'gate_proj.weight' in unpacked_state_dict and 'up_proj.weight' in unpacked_state_dict:
-            gate_proj_weight = unpacked_state_dict['gate_proj.weight']
-            up_proj_weight = unpacked_state_dict['up_proj.weight']
-            # Concatenate gate and up weights along output dimension (dim=0)
-            packed_state_dict['swiglu.w12.weight'] = torch.cat([gate_proj_weight, up_proj_weight], dim=0)
-
-        # Handle w1 and w3 -> w13 conversion for biases (if they exist)
-        if 'gate_proj.bias' in unpacked_state_dict and 'up_proj.bias' in unpacked_state_dict:
-            gate_proj_bias = unpacked_state_dict['gate_proj.bias']
-            up_proj_bias = unpacked_state_dict['up_proj.bias']
-            # Concatenate gate and up biases along dimension 0
-            packed_state_dict['swiglu.w12.bias'] = torch.cat([gate_proj_bias, up_proj_bias], dim=0)
-
-        # Copy down parameters as-is
-        if "down_proj.weight" in unpacked_state_dict:
-            packed_state_dict["swiglu.w3.weight"] = unpacked_state_dict['down_proj.weight']
-        if "down_proj.bias" in unpacked_state_dict:
-            packed_state_dict["swiglu.w3.bias"] = unpacked_state_dict['down_proj.bias']
-
-        for key in ['swiglu.w3.weight', 'swiglu.w3.bias']:
-            if key in unpacked_state_dict:
-                packed_state_dict[key] = unpacked_state_dict[key]
-
-        # Copy any other parameters that might exist
-        excluded_keys = [
-            'gate_proj.weight', 'gate_proj.bias',
-            'down_proj.weight', 'down_proj.bias',
-            'up_proj.weight', 'up_proj.bias',
-            'swiglu.w12.weight', 'swiglu.w12.bias',
-            'swiglu.w3.weight', 'swiglu.w3.bias',
-        ]
-        for key, value in unpacked_state_dict.items():
-            if key not in excluded_keys:
-                packed_state_dict[key] = value
-
-        return packed_state_dict
-
-    def load_state_dict(self, state_dict, strict=True):
-        """
-        Load state dict, handling both packed (w13) and unpacked (w1, w3) formats.
-        """
-        # Check if this is an unpacked state dict (has w1 and w3 instead of w13)
-        has_unpacked_gate_up = 'gate_proj.weight' in state_dict and 'up_proj.weight' in state_dict
-        has_packed_swiglu = 'swiglu.w12.weight' in state_dict
-
-        if has_unpacked_gate_up and not has_packed_swiglu:
-            state_dict = self._convert_unpacked_to_packed_state_dict(state_dict)
-
-        return super().load_state_dict(state_dict, strict=strict)
-
-    def state_dict(self, destination=None, prefix='', keep_vars=False, packed=False):
-        """
-        Return state dict in unpacked format by default for compatibility.
-        Set packed=True to get the internal packed format.
-        """
-        if packed:
-            # Return the actual packed state dict
-            return super().state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars)
-        else:
-            # Return unpacked format for compatibility
-            return self.get_unpacked_state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars)
-
-    def get_unpacked_state_dict(self, destination=None, prefix='', keep_vars=False):
-        """
-        Convert current packed state dict to unpacked format for compatibility.
-        """
-        # Get the actual packed state dict first
-        packed_state_dict = super().state_dict(destination=None, prefix='', keep_vars=keep_vars)
-
-        if destination is None:
-            destination = OrderedDict()
-
-        # Handle w13 -> w1 and w3 conversion for weights
-        if f'{prefix}swiglu.w12.weight' in packed_state_dict:
-            w13_weight = packed_state_dict[f'{prefix}swiglu.w12.weight']
-            hidden_dim = w13_weight.shape[0] // 2
-            w1_weight, w3_weight = torch.split(w13_weight, hidden_dim, dim=0)
-            destination[f'{prefix}gate_proj.weight'] = w1_weight if not keep_vars else w1_weight.detach().requires_grad_(w1_weight.requires_grad)
-            destination[f'{prefix}up_proj.weight'] = w3_weight if not keep_vars else w3_weight.detach().requires_grad_(w3_weight.requires_grad)
-
-        # Handle w13 -> w1 and w3 conversion for biases (if they exist)
-        if f'{prefix}swiglu.w12.bias' in packed_state_dict:
-            w13_bias = packed_state_dict[f'{prefix}swiglu.w12.bias']
-            hidden_dim = w13_bias.shape[0] // 2
-            w1_bias, w3_bias = torch.split(w13_bias, hidden_dim, dim=0)
-            destination[f'{prefix}gate_proj.bias'] = w1_bias if not keep_vars else w1_bias.detach().requires_grad_(w1_bias.requires_grad)
-            destination[f'{prefix}up_proj.bias'] = w3_bias if not keep_vars else w3_bias.detach().requires_grad_(w3_bias.requires_grad)
-
-        # Copy w2 parameters as-is
-        for param_name in ['weight', 'bias']:
-            key = f'{prefix}swiglu.w3.{param_name}'
-            if key in packed_state_dict:
-                destination[f'{prefix}down_proj.{param_name}'] = packed_state_dict[key]
-
-        # Copy any other parameters
-        excluded_prefixes = [f'{prefix}swiglu.w12.', f'{prefix}swiglu.w3.']
-        for key, value in packed_state_dict.items():
-            if not any(key.startswith(excluded_prefix) for excluded_prefix in excluded_prefixes) and key not in destination:
-                destination[key] = value
-
-        return destination
-
-    def forward(self, x):
-        return self.swiglu(x)