make sure to validate the config before normalizing so defaults get set (#2554)

* make sure to validate the config before normalizing so defaults get set

* validation not needed for particular test

* remove duplicate validations

* set qlora correctly

.coveragerc

@@ -0,0 +1,14 @@
+[run]
+source = axolotl
+omit =
+    */tests/*
+    setup.py
+
+[report]
+exclude_lines =
+    pragma: no cover
+    def __repr__
+    raise NotImplementedError
+    if __name__ == .__main__.:
+    pass
+    raise ImportError

.github/workflows/base.yml

@@ -46,6 +46,18 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.7.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "128"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.7.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""

.github/workflows/main.yml

@@ -29,8 +29,13 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras:
+            axolotl_extras: vllm
             is_latest: true
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras: vllm
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -93,6 +98,11 @@ jobs:
             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:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -138,7 +148,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.4.1
+            pytorch: 2.6.0
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:

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

@@ -45,6 +45,13 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
+            num_gpus: 2
+            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -67,6 +74,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.multigpu

.github/workflows/tests-nightly.yml

@@ -147,6 +147,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests

.github/workflows/tests.yml

@@ -49,7 +49,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
+        pytorch_version: ["2.4.1", "2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -102,9 +102,17 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v tests/patched/
-          pytest -v tests/cli/
+          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
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
 
       - name: cleanup pip cache
         run: |
@@ -234,6 +242,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
@@ -261,6 +270,12 @@ jobs:
             pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras: vllm
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -281,6 +296,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests

README.md

@@ -9,6 +9,7 @@
 <p align="center">
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests">
+    <a href="https://codecov.io/gh/axolotl-ai-cloud/axolotl"><img src="https://codecov.io/gh/axolotl-ai-cloud/axolotl/branch/main/graph/badge.svg" alt="codecov"></a>
     <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a>
     <br/>
     <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&style=flat-square" alt="contributors" style="height: 20px;"></a>

cicd/cicd.sh

@@ -3,10 +3,53 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli /workspace/axolotl/tests/
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
-pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
-pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
-pytest -v --durations=10 /workspace/axolotl/tests/cli
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ --ignore=tests/cli /workspace/axolotl/tests/e2e/
+# Run unit tests with initial coverage report
+pytest -v --durations=10 -n8 \
+  --ignore=tests/e2e/ \
+  --ignore=tests/patched/ \
+  --ignore=tests/cli \
+  /workspace/axolotl/tests/ \
+  --cov=axolotl
+
+# Run lora kernels tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/patched/lora_kernels \
+  --cov=axolotl \
+  --cov-append
+
+# Run patched tests excluding lora kernels with coverage append
+pytest -v --durations=10 \
+  --ignore=tests/e2e/patched/lora_kernels \
+  /workspace/axolotl/tests/e2e/patched \
+  --cov=axolotl \
+  --cov-append
+
+# Run solo tests with coverage append
+pytest -v --durations=10 -n1 \
+  /workspace/axolotl/tests/e2e/solo/ \
+  --cov=axolotl \
+  --cov-append
+
+# Run integration tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/integrations/ \
+  --cov=axolotl \
+  --cov-append
+
+pytest -v --durations=10 /workspace/axolotl/tests/cli \
+  --cov=axolotl \
+  --cov-append
+
+# Run remaining e2e tests with coverage append and final report
+pytest -v --durations=10 \
+  --ignore=tests/e2e/solo/ \
+  --ignore=tests/e2e/patched/ \
+  --ignore=tests/e2e/multigpu/ \
+  --ignore=tests/e2e/integrations/ \
+  --ignore=tests/cli \
+  /workspace/axolotl/tests/e2e/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:e2e-coverage.xml
+
+codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION}

cicd/e2e_tests.py

@@ -28,6 +28,7 @@ df_args = {
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
+    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
 }
 

cicd/multigpu.py

@@ -29,6 +29,7 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "121"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
 }
 

cicd/multigpu.sh

@@ -1,6 +1,23 @@
 #!/bin/bash
 set -e
 
-# only run one test at a time so as not to OOM the GPU
-pytest -v  --durations=10 -n2 /workspace/axolotl/tests/e2e/multigpu/ --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
-pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/solo/
+# Only run two tests at a time to avoid OOM on GPU (with coverage collection)
+pytest -v -n2 \
+  --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
+  --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
+  /workspace/axolotl/tests/e2e/multigpu/ \
+  --cov=axolotl
+
+# Run solo tests with coverage append
+pytest -v --durations=10 -n1 \
+  /workspace/axolotl/tests/e2e/multigpu/solo/ \
+  --cov=axolotl \
+  --cov-append
+
+pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/patched/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:multigpu-coverage.xml
+
+# Upload coverage to Codecov
+codecov upload-process -t $CODECOV_TOKEN -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION}

codecov.yml

@@ -0,0 +1,54 @@
+codecov:
+  require_ci_to_pass: yes
+
+coverage:
+  precision: 2
+  round: down
+  range: "70...100"
+  status:
+    project:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+    patch:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+
+parsers:
+  gcov:
+    branch_detection:
+      conditional: yes
+      loop: yes
+      method: no
+      macro: no
+
+comment:
+  layout: "reach,diff,flags,files,footer"
+  behavior: default
+  require_changes: no
+  require_base: no
+  require_head: yes
+
+github_checks:
+  annotations: false

docker/Dockerfile-base

@@ -37,3 +37,7 @@ RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10
+
+RUN if [ "$PYTORCH_VERSION" = "2.7.0" ] ; then \
+        pip3 install flash-attn==2.7.4.post1; \
+    fi

docs/cli.qmd

@@ -199,6 +199,17 @@ output_dir: # Directory to save evaluation results
 
 See [LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) for more details.
 
+### delinearize-llama4
+
+Delinearizes a Llama 4 linearized model into a regular HuggingFace Llama 4 model. This only works with the non-quantized linearized model.
+
+```bash
+axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
+```
+
+This would be necessary to use with other frameworks. If you have an adapter, merge it with the non-quantized linearized model before delinearizing.
+
+
 ## Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:

docs/config.qmd

@@ -693,6 +693,9 @@ sequence_parallel_degree:
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 # Must evenly divide the number of KV heads in your model.
 heads_k_stride: 1
+# One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3"
+# in the sample packing case, and "batch_ring" in the non-sample packing case.
+ring_attn_func:
 
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:

docs/installation.qmd

@@ -19,6 +19,12 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 ## Installation Methods {#sec-installation-methods}
 
+::: {.callout-important}
+Please make sure to have Pytorch installed before installing Axolotl in your local environment.
+
+Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
+:::
+
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}

docs/sequence_parallelism.qmd

@@ -27,6 +27,9 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
+ring_attn_func:
 ```
 
 The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:

examples/glm4/qlora-32b.yaml

@@ -0,0 +1,62 @@
+base_model: THUDM/GLM-4-32B-0414
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_4bit: true
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+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_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/llama-4/README.md

@@ -1,16 +1,36 @@
 # Llama 4 by Meta AI
 
+## Flash Attention vs Flex Attention
+
+While Flash Attention to support is "enabled" for Llama-4, the upstream implementation is not correct and usage of Flex Attention is recommended.
+
 ## Available Examples
 
 ### Llama 4 Scout 17Bx16Experts (109B)
-- [Multi-Modal/Vision QLoRA w/ FSDP1](./scout-vision-qlora-fsdp.yaml)
-- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100.yaml)
-- [Text Multi GPU QLoRA w/ FSDP1](./scout-qlora-fsdp1.yaml)
 
-Our Single H100 implementation for Llama 4 Scout uses only 68.5GB VRAM for post-training with 4k context length @ 546 tokens/second. [WandB logs here](https://wandb.ai/axolotl-ai/llama4-sft/runs/zic56rhd)
+Flex Attention
+- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100-flex.yaml)
+- [Text Multi GPU QLoRA w/ FSDP2](./scout-qlora-flexattn-fsdp2.yaml)
+
+[//]: # (Flash Attention (Do not use))
+
+[//]: # (- [Multi-Modal/Vision QLoRA w/ FSDP1](./scout-vision-qlora-fsdp.yaml))
+
+[//]: # (- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100.yaml))
+
+[//]: # (- [Text Multi GPU QLoRA w/ FSDP1](./scout-qlora-fsdp1.yaml))
+
+Our Single H100 implementation for Llama 4 Scout uses only 64.5GB VRAM for post-training with 4k context length @ 519 tokens/second. [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/wpie7dkj)
+Multi-GPU (4xH100) for Llama 4 Scout uses 62.8GB VRAM/GPU @ 4k contenxt length @ 280tps/gpu, [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/2lkezdj8)
 
 ### Llama 4 Maverick 17Bx128Experts (400B)
 
-- [Text Multi GPU QLoRA w/FSDP1](./maverick-qlora-fsdp1.yaml)
+Coming Soon
 
-Our 4xH100 implementation for Llama 4 Maverick uses 79.5GB VRAM/GPU for post-training with 4k context length @ 206 tokens/second. [WandB logs here.](https://wandb.ai/axolotl-ai/llama-sft/runs/siyvwuxc?nw=nwuserwinglian)
+## Delinearized Llama 4 Models
+
+We provide a script to delinearize Llama 4 linearized models into regular HuggingFace Llama 4 models.
+
+```bash
+axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
+```

examples/llama-4/scout-qlora-flexattn-fsdp2.yaml

@@ -0,0 +1,86 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head
+  # - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-qlora-single-h100-flex.yaml

@@ -0,0 +1,85 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+cut_cross_entropy: true
+
+llama4_linearized_experts: true  # needed with custom linearized experts model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$  # optionally train the moe experts
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head  # needed if modifying vocabulary
+  # - embed_tokens
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096  # up to 8k will work on a single H100
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+torch_compile: true
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+logging_steps: 1
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-vision-qlora-fsdp2-flex.yaml

@@ -0,0 +1,89 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+processor_type: Llama4Processor
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+sequence_len: 4096
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true  # use Axolotl's customized model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - vision_adapter.mlp.fc1
+  - vision_adapter.mlp.fc2
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

requirements-dev.txt

@@ -1,6 +1,6 @@
-pre-commit
 black
 mypy
+pre-commit
 types-requests
 quartodoc
 jupyter

requirements-tests.txt

@@ -1,5 +1,8 @@
+codecov
+codecov-cli
 pytest
-pytest-xdist
+pytest-cov
 pytest-retry
 pytest-sugar
+pytest-xdist
 tbparse

requirements.txt

@@ -6,7 +6,7 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.6
+liger-kernel==0.5.8
 # END section
 
 packaging==23.2
@@ -19,6 +19,7 @@ datasets==3.5.0
 deepspeed>=0.15.4
 trl==0.16.1
 hf_xet==1.0.0
+hqq==0.2.5
 
 optimum==1.16.2
 hf_transfer

scripts/cutcrossentropy_install.py

@@ -25,5 +25,5 @@ if cce_spec:
 
 print(
     UNINSTALL_PREFIX
-    + 'pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"'
+    + 'pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"'
 )

setup.py

@@ -51,7 +51,7 @@ def parse_requirements(extras_require_map):
             try:
                 torch_version = version("torch")
             except PackageNotFoundError:
-                torch_version = "2.5.1"
+                torch_version = "2.6.0"  # default to torch 2.6
             _install_requires.append(f"torch=={torch_version}")
 
             version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
@@ -64,10 +64,16 @@ def parse_requirements(extras_require_map):
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 6):
+            if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append("xformers==0.0.29.post2")
-                extras_require_map["vllm"] = ["vllm==0.8.1"]
+                # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
+                extras_require_map["vllm"] = ["vllm==0.8.3"]
+            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==0.8.3"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:

src/axolotl/cli/args.py

@@ -39,16 +39,16 @@ class TrainerCliArgs:
 class VllmServeCliArgs:
     """Dataclass with CLI arguments for `axolotl vllm-serve` command."""
 
-    tensor_parallel_size: int = field(
-        default=1,
+    tensor_parallel_size: Optional[int] = field(
+        default=None,
         metadata={"help": "Number of tensor parallel workers to use."},
     )
-    host: str = field(
-        default="0.0.0.0",  # nosec B104
+    host: Optional[str] = field(
+        default=None,  # nosec B104
         metadata={"help": "Host address to run the server on."},
     )
-    port: int = field(
-        default=8000,
+    port: Optional[int] = field(
+        default=None,
         metadata={"help": "Port to run the server on."},
     )
     gpu_memory_utilization: Optional[float] = field(

src/axolotl/cli/delinearize_llama4.py

@@ -0,0 +1,156 @@
+"""
+CLI tool to delinearize quantized/Linearized Llama-4 models.
+"""
+
+import os
+from pathlib import Path
+from typing import Generator, Union
+
+import fire
+import torch
+from accelerate import init_empty_weights
+from dotenv import load_dotenv
+from transformers import AutoProcessor
+
+
+def iter_convert_patched_to_hf(model_state_dict, num_experts) -> Generator:
+    keys = list(model_state_dict.keys())
+    for key in keys:
+        if ".feed_forward.experts." not in key:
+            yield key, model_state_dict[key]
+        if ".feed_forward.experts.gate_projs" in key:
+            # gate gets fused with up so skip the yield on this and we'll fuse it when asking for the up
+            continue
+        if ".feed_forward.experts.up_projs" in key:
+            if ".feed_forward.experts.up_projs.0." in key:
+                # handle the re-shape and fusing of gate and up, and conversion from linear to parameter
+                prefix = key.split(".up_projs.0.")[0]
+                key = f"{prefix}.gate_up_proj"
+                # grab all the up_projs and gate_projs across all experts
+                gate_stacked = torch.stack(
+                    [
+                        model_state_dict[
+                            f"{prefix}.gate_projs.{expert_idx}.weight"
+                        ].transpose(0, 1)
+                        for expert_idx in range(num_experts)
+                    ]
+                )
+                up_stacked = torch.stack(
+                    [
+                        model_state_dict[
+                            f"{prefix}.up_projs.{expert_idx}.weight"
+                        ].transpose(0, 1)
+                        for expert_idx in range(num_experts)
+                    ]
+                )
+                gate_up_proj = torch.cat((gate_stacked, up_stacked), dim=-1)
+                del gate_stacked, up_stacked
+                yield key, gate_up_proj
+            else:
+                del model_state_dict[key]
+                continue
+        if ".feed_forward.experts.down_projs" in key:
+            if ".feed_forward.experts.down_projs.0." in key:
+                # handle the re-shape and fusing of gate and up, and conversion from linear to parameter
+                prefix = key.split(".down_projs.0.")[0]
+                key = f"{prefix}.down_proj"
+                # grab all the down_projs across all experts
+                down_stacked = torch.stack(
+                    [
+                        model_state_dict[
+                            f"{prefix}.down_projs.{expert_idx}.weight"
+                        ].transpose(0, 1)
+                        for expert_idx in range(num_experts)
+                    ]
+                )
+                yield key, down_stacked
+            else:
+                del model_state_dict[key]
+                continue
+
+
+def do_cli(model: Union[Path, str], output: Union[Path, str]) -> None:
+    """
+    Convert a patched HF format Llama4 model (with separated projections)
+    back to the original HF format (with fused projections).
+
+    Args:
+        model: Path to the patched HF model
+        output: Path to save the converted model
+    """
+    print(f"Loading model from {model}")
+    from axolotl.monkeypatch.models.llama4.modeling import (
+        patch_llama4_linearized_modeling,
+    )
+
+    unpatch_llama4 = patch_llama4_linearized_modeling()
+    from transformers import Llama4ForConditionalGeneration
+
+    model_ = Llama4ForConditionalGeneration.from_pretrained(
+        model, torch_dtype=torch.bfloat16
+    )
+    processor = AutoProcessor.from_pretrained(model)
+    processor.save_pretrained(output)
+
+    device = model_.device.type
+    if device == "cuda":
+        print(
+            f"peak memory allocated: {torch.cuda.max_memory_allocated() / 1024**2} MB"
+        )
+        print(f"peak memory reserved: {torch.cuda.max_memory_reserved() / 1024**2} MB")
+    model_config = model_.config
+    config = model_.config.get_text_config()
+
+    # Get key dimensions from the config
+    hidden_size = config.hidden_size
+    intermediate_size = config.intermediate_size
+    num_experts = config.num_local_experts
+
+    print(
+        f"Model dimensions: hidden_size={hidden_size}, intermediate_size={intermediate_size}, num_experts={num_experts}"
+    )
+
+    # Create output directory if it doesn't exist
+    os.makedirs(output, exist_ok=True)
+
+    # Get state dict
+    state_dict = model_.state_dict()
+    del model_
+
+    # Create a new state dict for the converted model
+    converted_state_dict = {}
+
+    # First, copy all keys that don't need modification
+    for key, value in iter_convert_patched_to_hf(state_dict, num_experts):
+        converted_state_dict[key] = value
+
+    del state_dict
+    if device == "cuda":
+        torch.cuda.empty_cache()
+        print("State dict converted.")
+        print(
+            f"peak memory allocated: {torch.cuda.max_memory_allocated() / 1024**2} MB"
+        )
+        print(f"peak memory reserved: {torch.cuda.max_memory_reserved() / 1024**2} MB")
+    # Ideally re-load the model import to load the converted state dict
+    # Save the converted model
+    with init_empty_weights():
+        unpatch_llama4()
+        model_ = Llama4ForConditionalGeneration(model_config)
+
+    if device == "cuda":
+        print("State dict loaded into model.")
+        print(
+            f"peak memory allocated: {torch.cuda.max_memory_allocated() / 1024**2} MB"
+        )
+        print(f"peak memory reserved: {torch.cuda.max_memory_reserved() / 1024**2} MB")
+    model_.load_state_dict(converted_state_dict, strict=False, assign=True)
+    print(f"Saving converted model to {output}...")
+    model_.save_pretrained(output)
+
+    print(f"Model successfully converted and saved to {output}")
+
+
+if __name__ == "__main__":
+    load_dotenv()
+    fire.Fire(do_cli)

src/axolotl/cli/main.py

@@ -330,6 +330,15 @@ def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
     do_vllm_serve(config, cli_args)
 
 
+@cli.command()
+@click.argument("model", type=click.Path(exists=True, path_type=str))
+@click.argument("output", type=click.Path(exists=False, path_type=str))
+def delinearize_llama4(model: str, output: str) -> None:
+    from axolotl.cli.delinearize_llama4 import do_cli as do_delinearize_llama4
+
+    do_delinearize_llama4(model, output)
+
+
 cli.add_command(lm_eval)
 
 

src/axolotl/cli/merge_lora.py

@@ -40,6 +40,7 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         LOG.warning("Error raised: %s", e)
 
     model.generation_config.do_sample = True
+    model.config.use_cache = True
 
     if cfg.local_rank == 0:
         LOG.info(f"Saving merged model to: {str(Path(cfg.output_dir) / 'merged')}...")

src/axolotl/core/trainer_builder.py

@@ -776,6 +776,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["sequence_parallel_degree"] = (
             self.cfg.sequence_parallel_degree
         )
+        training_arguments_kwargs["ring_attn_func"] = self.cfg.ring_attn_func
 
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
@@ -933,6 +934,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         kwargs["return_tensors"] = "pt"
         if issubclass(collator, DataCollatorForSeq2Seq):
             kwargs["sequence_parallel_degree"] = training_args.sequence_parallel_degree
+            kwargs["ring_attn_func"] = training_args.ring_attn_func
 
         return collator(
             *collator_args,
@@ -1038,9 +1040,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.dataset_processes:
             training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
 
-        if (self.cfg.trl and self.cfg.trl.beta) or self.cfg.rl_beta:
-            training_args_kwargs["beta"] = self.cfg.trl.beta or self.cfg.rl_beta
-        if self.cfg.orpo_alpha:
+        if self.cfg.trl and self.cfg.trl.beta is not None:
+            training_args_kwargs["beta"] = self.cfg.trl.beta
+        elif self.cfg.rl_beta is not None:
+            training_args_kwargs["beta"] = self.cfg.rl_beta
+        elif self.cfg.orpo_alpha is not None:
             # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
             training_args_kwargs["beta"] = self.cfg.orpo_alpha
 

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

@@ -40,8 +40,8 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:

src/axolotl/core/training_args.py

@@ -9,6 +9,8 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
 
 @dataclass
 class AxolotlTrainingMixins:
@@ -218,6 +220,12 @@ class AxolotlTrainingMixins:
         default=1,
         metadata={"help": "The number of workers to use in sequence parallelism"},
     )
+    ring_attn_func: Optional[RingAttnFunc] = field(
+        default=None,
+        metadata={
+            "help": "The ring-flash-attn function to use in sequence parallelism"
+        },
+    )
 
     # multi-modal section
 

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -12,12 +12,14 @@ See https://github.com/apple/ml-cross-entropy
 
 Run the following command to install `cut_cross_entropy[transformers]` if you don't have it already.
 
+- If you are in dev environment
 ```bash
-# if you are in dev environment
 python scripts/cutcrossentropy_install.py | sh
+```
 
-# if you are not in dev environment
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"
+- If you are installing from pip
+```bash
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"
 ```
 
 ## Usage
@@ -45,6 +47,8 @@ cut_cross_entropy: true
 - qwen2
 - cohere
 - cohere2
+- glm
+- glm4
 
 ## Citation
 

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -33,7 +33,7 @@ LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy")
 
 _CCE_INSTALL_MESSAGE = (
     "Please install cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"`'
 )
 
 

src/axolotl/integrations/cut_cross_entropy/monkeypatch/glm4.py

@@ -0,0 +1,57 @@
+"""GLM 4 patch. GLM family inherits from Llama."""
+
+from types import MethodType
+
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+
+
+def patch_glm(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import cut_cross_entropy.transformers.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from cut_cross_entropy.transformers.llama import cce_forward
+    from transformers.models.glm import modeling_glm
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_glm.GlmForCausalLM
+        ), f"Expected a GlmForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_glm.GlmForCausalLM.forward = cce_forward
+    return None
+
+
+def patch_glm4(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import cut_cross_entropy.transformers.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from cut_cross_entropy.transformers.llama import cce_forward
+    from transformers.models.glm4 import modeling_glm4
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_glm4.Glm4ForCausalLM
+        ), f"Expected a Glm4ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_glm4.Glm4ForCausalLM.forward = cce_forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/llama4.py

@@ -165,7 +165,7 @@ def cce_forward(
 )
 def cce_forward_multimodal(
     self,
-    input_ids: torch.LongTensor | None = None,
+    input_ids: torch.LongTensor | None = None,  # type: ignore
     pixel_values: torch.FloatTensor | None = None,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
@@ -254,7 +254,7 @@ def cce_forward_multimodal(
         )
 
     if inputs_embeds is None:
-        inputs_embeds = self.get_input_embeddings()(input_ids)
+        inputs_embeds = self.get_input_embeddings()(input_ids)  # type: ignore
 
     if pixel_values is not None:
         image_features = self.get_image_features(
@@ -263,13 +263,13 @@ def cce_forward_multimodal(
             vision_feature_select_strategy=vision_feature_select_strategy,
             image_sizes=image_sizes,
         )
-        original_inputs_embeds_shape = inputs_embeds.shape
+        original_inputs_embeds_shape = inputs_embeds.shape  # type: ignore
 
         vision_flat = image_features.view(-1, image_features.size(-1))
         projected_vision_flat = self.multi_modal_projector(vision_flat)
 
         special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
-        final_mask = special_image_mask.to(inputs_embeds.device)
+        final_mask = special_image_mask.to(inputs_embeds.device)  # type: ignore
         inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1))  # type: ignore
 
         final_mask_1d = final_mask[..., 0].reshape(-1)

src/axolotl/integrations/cut_cross_entropy/monkeypatch/patch.py

@@ -20,6 +20,10 @@ from axolotl.integrations.cut_cross_entropy.monkeypatch.gemma3 import (
     patch_gemma3,
     patch_gemma3_text,
 )
+from axolotl.integrations.cut_cross_entropy.monkeypatch.glm4 import (
+    patch_glm,
+    patch_glm4,
+)
 from axolotl.integrations.cut_cross_entropy.monkeypatch.llama4 import (
     patch_llama4,
     patch_llama4_text,
@@ -45,6 +49,8 @@ CUT_CROSS_ENTROPY_MODEL_MAPPING = {
     "qwen2": patch_qwen2,
     "cohere": patch_cohere,
     "cohere2": patch_cohere2,
+    "glm": patch_glm,
+    "glm4": patch_glm4,
 }
 
 

src/axolotl/integrations/liger/README.md

@@ -25,7 +25,7 @@ liger_fused_linear_cross_entropy: true
 - deepseek_v2
 - gemma
 - gemma2
-- gemma3 (partial support, no support for FLCE yet)
+- gemma3
 - granite
 - jamba
 - llama

src/axolotl/integrations/liger/__init__.py

@@ -21,7 +21,6 @@ It is designed to be performant, correct, and light-weight.
 import inspect
 import logging
 import sys
-from functools import partial
 
 from axolotl.integrations.base import BasePlugin
 
@@ -55,7 +54,6 @@ class LigerPlugin(BasePlugin):
             )
         from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
         from liger_kernel.transformers.functional import liger_cross_entropy
-        from liger_kernel.transformers.geglu import LigerGEGLUMLP
         from liger_kernel.transformers.layer_norm import LigerLayerNorm
         from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
         from liger_kernel.transformers.rms_norm import LigerRMSNorm
@@ -141,38 +139,6 @@ class LigerPlugin(BasePlugin):
                 modeling_mod.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_mod.DeepseekV2ForCausalLM.forward = deepseekv2_lce_forward
-        elif cfg.model_config_type in ["gemma3", "gemma3_text"]:
-            from transformers.models.gemma3 import modeling_gemma3
-
-            if cfg.liger_rope:
-                modeling_gemma3.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-
-                def _liger_rms_norm_wrapper(dim, **kwargs):
-                    "Convert 'dim' keyword to 'hidden_size' to pass to LigerRMSNorm"
-                    return LigerRMSNorm(hidden_size=dim, **kwargs)
-
-                modeling_gemma3.Gemma3RMSNorm = partial(
-                    _liger_rms_norm_wrapper,
-                    offset=1.0,
-                    casting_mode="gemma",
-                    init_fn="zeros",
-                    in_place=False,
-                )
-            if cfg.liger_glu_activation:
-                modeling_gemma3.Gemma3MLP = LigerGEGLUMLP
-            if cfg.liger_layer_norm:
-                modeling_gemma3.nn.LayerNorm = LigerLayerNorm
-
-            if cfg.liger_cross_entropy:
-                from transformers.loss.loss_utils import nn
-
-                nn.functional.cross_entropy = liger_cross_entropy
-
-            if cfg.liger_fused_linear_cross_entropy:
-                raise NotImplementedError(
-                    "Fused linear cross entropy is not yet supported for Gemma3."
-                )
         elif cfg.model_config_type == "llama4":
             from axolotl.integrations.liger.models.llama4 import (
                 apply_liger_kernel_to_llama4,

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -49,7 +49,7 @@ def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dic
             )
             sharded_sd[param_name] = sharded_tensor
 
-    model.load_state_dict(sharded_sd)
+    model.load_state_dict(sharded_sd, assign=True)
 
 
 def patch_accelerate_fsdp_utils():

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -7,12 +7,11 @@ import torch
 import transformers
 
 
-def patch_flex_wrapper():
+def patch_flex_wrapper(**flex_attn_compile_kwargs):
     # TODO remove this patch when transformers#37285 is merged and in a release
     is_torch_2_6 = torch.__version__.startswith("2.6")
-    is_transformers_below_4_51 = transformers.__version__ < "4.51.0"
 
-    if not (is_torch_2_6 and is_transformers_below_4_51):
+    if not is_torch_2_6:
         return
 
     from torch.nn.attention.flex_attention import flex_attention
@@ -32,17 +31,24 @@ def patch_flex_wrapper():
                 cls._instance = super().__new__(cls)
             return cls._instance
 
+        @classmethod
+        def del_singleton(cls):
+            cls._instance = None
+
         @torch.compiler.disable(recursive=False)
-        def __init__(self):
+        def __init__(self, training):
             """
             Initialize or update the singleton instance.
             """
-            if not self._is_flex_compiled:
+            self.training = None
+            if not self._is_flex_compiled or training != self.training:
+                # In PyTorch 2.6.0, there's a known issue with flex attention compilation which may
+                # 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
                 self._compiled_flex_attention = torch.compile(
                     flex_attention,
-                    dynamic=False,
-                    mode="max-autotune-no-cudagraphs",
-                    fullgraph=True,
+                    **flex_attn_compile_kwargs,
                 )
                 self._is_flex_compiled = True
 
@@ -50,15 +56,22 @@ def patch_flex_wrapper():
             return self._compiled_flex_attention
 
     transformers.integrations.flex_attention.WrappedFlexAttention = WrappedFlexAttention
+    setattr(
+        sys.modules["transformers.integrations.flex_attention"],
+        "WrappedFlexAttention",
+        WrappedFlexAttention,
+    )
 
 
 def patch_flex_make_mask():
     is_torch_2_6 = torch.__version__.startswith("2.6")
-    is_transformers_eq_4_51 = transformers.__version__ == "4.51.0"
 
-    if not (is_torch_2_6 and is_transformers_eq_4_51):
+    if not is_torch_2_6:
         return
 
+    from torch.nn.attention.flex_attention import (
+        _DEFAULT_SPARSE_BLOCK_SIZE as flex_default_block_size,
+    )
     from torch.nn.attention.flex_attention import (
         BlockMask,
     )
@@ -104,14 +117,16 @@ def patch_flex_make_mask():
         if not query_length:
             query_length = total_seq_len
         attention_mask_2d = torch.nn.functional.pad(
-            attention_mask_2d, value=0, pad=(0, key_length)
+            attention_mask_2d,
+            value=0,
+            pad=(0, abs(total_seq_len - max(key_length, flex_default_block_size))),
         )
         device = attention_mask_2d.device
         document_ids = attention_mask_2d.clone()
 
         if attention_chunk_size is not None:
             # we create an arange, then we just // by chunk size to get [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3]
-            document_ids = (document_ids.fill_(1).cumsum(-1) - 1) // (
+            chunk_idxs = (document_ids.clone().fill_(1).cumsum(-1) - 1) // (
                 attention_chunk_size
             )
 
@@ -138,6 +153,18 @@ def patch_flex_make_mask():
             final_mask = causal_mask & padding_mask & document_mask
             return final_mask
 
+        def chunk_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
+            """
+            Combines the chunk mask with the causal mask for chunked attention.
+            """
+            chunk_mask = chunk_idxs[batch_idx, q_idx] == chunk_idxs[batch_idx, kv_idx]
+            causal_doc_mask = causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx)
+            return chunk_mask & causal_doc_mask
+
+        mask_mod_maybe_combined = (
+            causal_mask_mod if attention_chunk_size is None else chunk_causal_mask_mod
+        )
+
         if offsets is not None:
             q_offset = offsets[0]
             kv_offset = offsets[1]
@@ -145,10 +172,10 @@ def patch_flex_make_mask():
             def mask_mod(batch_idx, head_idx, q_idx, kv_idx):
                 offset_q = q_idx + q_offset
                 offset_kv = kv_idx + kv_offset
-                return causal_mask_mod(batch_idx, head_idx, offset_q, offset_kv)
+                return mask_mod_maybe_combined(batch_idx, head_idx, offset_q, offset_kv)
 
         else:
-            mask_mod = causal_mask_mod
+            mask_mod = mask_mod_maybe_combined
         return create_block_causal_mask_flex(
             mask_mod=mask_mod,
             B=batch_size,
@@ -160,11 +187,16 @@ def patch_flex_make_mask():
         )
 
     for n in tuple(sys.modules):
-        if ".modeling_" in n and "llama4" not in n:
+        if ".modeling_" in n:
             if hasattr(sys.modules[n], "make_flex_block_causal_mask"):
                 sys.modules[n].make_flex_block_causal_mask = (
                     patched_make_flex_block_causal_mask
                 )
+                setattr(
+                    sys.modules[n],
+                    "make_flex_block_causal_mask",
+                    patched_make_flex_block_causal_mask,
+                )
 
     transformers.integrations.flex_attention.make_flex_block_causal_mask = (
         patched_make_flex_block_causal_mask

src/axolotl/monkeypatch/attention/ring_attn/__init__.py

@@ -0,0 +1,12 @@
+"""Init for ring attention monkeypatch module"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .patch import (
+    RingAttnFunc,
+    get_ring_attn_group,
+    register_ring_attn,
+    set_ring_attn_group,
+    update_ring_attn_params,
+)

src/axolotl/monkeypatch/attention/ring_attn/adapters/batch.py

@@ -0,0 +1,192 @@
+"""
+HuggingFace flash attention adapter for basic ring attention (batch API).
+
+Inspired by
+https://github.com/zhuzilin/ring-flash-attention/blob/ce9fd3935ca0e5f0592bb0826cbed18ec69da729/ring_flash_attn/adapters/hf_adapter.py.
+Our implementation closely follows the structure of that module, but we've minified it
+somewhat to support only the latest versions of transformers.
+"""
+
+# pylint: disable=protected-access,cyclic-import
+
+import os
+from typing import Callable
+
+import torch
+import torch.distributed as dist
+import transformers
+import transformers.modeling_flash_attention_utils
+from ring_flash_attn import (
+    ring_flash_attn_func,
+    stripe_flash_attn_func,
+    zigzag_ring_flash_attn_func,
+)
+from ring_flash_attn.adapters.hf_adapter import check_params
+from transformers.modeling_flash_attention_utils import (
+    _flash_supports_window_size,
+    is_flash_attn_greater_or_equal,
+)
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
+RING_ATTN_FUNC_MAPPING = {
+    RingAttnFunc.BATCH_RING: ring_flash_attn_func,
+    RingAttnFunc.BATCH_ZIGZAG: zigzag_ring_flash_attn_func,
+    RingAttnFunc.BATCH_STRIPE: stripe_flash_attn_func,
+}
+
+
+def create_flash_attn_forward(
+    process_group: dist.ProcessGroup, ring_attn_func: RingAttnFunc
+) -> Callable:
+    """
+    Create a ring flash attention forward function compatible with HuggingFace's
+    interface.
+
+    Args:
+        process_group: A PyTorch distributed process group.
+        ring_attn_func: Function from `ring_flash_attention` to replace HF flash
+            attention with.
+
+    Returns:
+        A function that implements the ring flash attention forward pass with the
+            signature expected by HuggingFace Transformers.
+    """
+
+    # transformers 4.48+
+    # pylint: disable=unused-argument
+    def _flash_attention_forward(
+        query_states: torch.Tensor,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        query_length: int,
+        is_causal: bool,
+        dropout: float = 0.0,
+        position_ids: torch.Tensor | None = None,
+        softmax_scale: float | None = None,
+        sliding_window: int | None = None,
+        use_top_left_mask: bool = False,
+        softcap: float | None = None,
+        deterministic: bool = None,
+        cu_seq_lens_q: torch.LongTensor | None = None,
+        cu_seq_lens_k: torch.LongTensor | None = None,
+        max_length_q: int | None = None,
+        max_length_k: int | None = None,
+        target_dtype: torch.dtype | None = None,
+        **kwargs,
+    ):
+        """
+        Calls the forward method of Ring Flash Attention.
+
+        Args:
+            query_states: Tensor containing the query vectors.
+            key_states: Tensor containing the key vectors.
+            value_states: Tensor containing the value vectors.
+            attention_mask: Not used in this implementation.
+            query_length: Integer representing the length of the query sequence.
+            is_causal: Boolean indicating whether to apply a causal mask to the attention.
+            dropout: Float representing the dropout probability. Default is 0.0.
+            position_ids: Not used in this implementation.
+            softmax_scale: Optional float value for the softmax scaling factor. Default is None.
+            sliding_window: Optional integer defining the size of the sliding attention window.
+                Default is None.
+            use_top_left_mask: Boolean indicating whether to use a top-left mask for the attention.
+                Default is False.
+            softcap: Not used in this implementation.
+            deterministic: Optional boolean to enforce deterministic computation. Default is None.
+            cu_seq_lens_q: Not used in this implementation.
+            cu_seq_lens_k: Not used in this implementation.
+            max_length_q: Not used in this implementation.
+            max_length_k: Not used in this implementation.
+            target_dtype: Not used in this implementation.
+            **kwargs: Additional keyword arguments. Not used in this implementation.
+
+        Returns:
+            torch.Tensor: The output of the attention mechanism, with shape
+                `[batch_size, query_length, num_heads, head_dim]`.
+        """
+        if not use_top_left_mask:
+            causal = is_causal
+        else:
+            causal = is_causal and query_length != 1
+
+        # Handle sliding window
+        use_sliding_windows = (
+            _flash_supports_window_size
+            and sliding_window is not None
+            and key_states.shape[1] > sliding_window
+        )
+        window_size = (
+            (sliding_window, sliding_window) if use_sliding_windows else (-1, -1)
+        )
+
+        # Handle deterministic mode
+        if is_flash_attn_greater_or_equal("2.4.1"):
+            if deterministic is None:
+                deterministic = (
+                    os.environ.get("FLASH_ATTENTION_DETERMINISTIC", "0") == "1"
+                )
+
+        # Call ring flash attention function
+        attn_output = RING_ATTN_FUNC_MAPPING[ring_attn_func](
+            query_states,
+            key_states,
+            value_states,
+            dropout_p=dropout,
+            softmax_scale=softmax_scale,
+            causal=causal,
+            window_size=window_size,
+            alibi_slopes=None,
+            deterministic=deterministic,
+            return_attn_probs=False,
+            group=process_group,
+        )
+
+        return attn_output
+
+    return _flash_attention_forward
+
+
+def substitute_hf_flash_attn(
+    process_group: dist.ProcessGroup, ring_attn_func: RingAttnFunc
+):
+    """
+    Substitute HuggingFace's flash attention implementation with ring-based implementation.
+
+    Args:
+        process_group: PyTorch distributed process group for communication.
+        ring_attn_func: Function from `ring_flash_attention` to replace HF flash
+            attention with.
+    """
+    try:
+        # Substitute flash attention
+        old_flash_attention_forward = (
+            transformers.modeling_flash_attention_utils._flash_attention_forward
+        )
+        new_flash_attention_forward = create_flash_attn_forward(
+            process_group=process_group, ring_attn_func=ring_attn_func
+        )
+
+        if check_params(old_flash_attention_forward, new_flash_attention_forward):
+            transformers.modeling_flash_attention_utils._flash_attention_forward = (
+                new_flash_attention_forward
+            )
+        else:
+            raise ValueError(
+                "The signature of the new flash attention forward function does not match the old one."
+            )
+    except Exception as exception:
+        raise ValueError(
+            f"The current transformer version {transformers.__version__} is not supported. "
+            "Please use pip install -U transformers to upgrade to the latest version. "
+            "If the code failed with the latest version, "
+            f"please file an issue."
+        ) from exception
+
+    # Register with ALL_ATTENTION_FUNCTIONS if available
+    if ALL_ATTENTION_FUNCTIONS is not None:
+        from ring_flash_attn.adapters.hf_adapter import flash_attention_forward
+
+        ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward

src/axolotl/monkeypatch/attention/ring_attn/patch.py

@@ -6,6 +6,8 @@ package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patc
 their sequence parallel version of Flash Attention 2.
 """
 
+from enum import Enum
+
 import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
@@ -16,6 +18,7 @@ from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 configure_logging()
 LOG = get_logger(__name__)
 
+
 RING_ATTN_GROUP = None
 
 
@@ -40,7 +43,22 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
-def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None):
+class RingAttnFunc(str, Enum):
+    """Enum class for supported `ring-flash-attn` implementations"""
+
+    # VARLEN_RING = "varlen_ring"
+    # VARLEN_ZIGZAG = "varlen_zigzag"
+    VARLEN_LLAMA3 = "varlen_llama3"
+    BATCH_RING = "batch_ring"
+    BATCH_ZIGZAG = "batch_zigzag"
+    BATCH_STRIPE = "batch_stripe"
+
+
+def register_ring_attn(
+    sequence_parallel_degree: int,
+    heads_k_stride: int | None,
+    ring_attn_func: RingAttnFunc | None,
+):
     """
     Create ring attention group and substitute flash attn with ring flash attn.
 
@@ -48,6 +66,9 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
         sequence_parallel_degree: Sequence parallelism factor.
         heads_k_stride: Sequence parallelism K head stride size. Passed
             through to `ring_flash_attn.substitute_hf_flash_attn`.
+        ring_attn_func: `ring_flash_attn` ring attention implemention. If sample
+            packing is enabled, it must be a `varlen` function; otherwise, it must be a
+            `batch` function.
     """
     if get_ring_attn_group() is not None:
         LOG.info("Ring attention already registered, exiting early...")
@@ -58,7 +79,9 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
         f"each sequence will be processed across {sequence_parallel_degree} GPUs"
     )
 
+    rank = dist.get_rank()
     world_size = dist.get_world_size()
+
     assert sequence_parallel_degree <= world_size, (
         f"sequence_parallel_degree ({sequence_parallel_degree}) "
         f"must be less than or equal to world_size ({world_size})"
@@ -68,10 +91,8 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
         f"must evenly divide world_size ({world_size})"
     )
 
-    # Detailed logging of group formation
-    rank = dist.get_rank()
+    # Assign ranks to sequence parallel groups
     group_assignments = {}
-
     for i in range(world_size // sequence_parallel_degree):
         ring_attn_ranks = list(
             range(
@@ -92,35 +113,37 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
     if rank == 0:
         LOG.info(f"Sequence parallel group assignments: {group_assignments}")
 
-    if heads_k_stride is None:
-        heads_k_stride = 1
+    if ring_attn_func is RingAttnFunc.VARLEN_LLAMA3:
+        from ring_flash_attn import substitute_hf_flash_attn
 
-    from ring_flash_attn import substitute_hf_flash_attn
+        substitute_hf_flash_attn(
+            process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
+        )
+    elif ring_attn_func in [
+        RingAttnFunc.BATCH_RING,
+        RingAttnFunc.BATCH_ZIGZAG,
+        RingAttnFunc.BATCH_STRIPE,
+    ]:
+        from axolotl.monkeypatch.attention.ring_attn.adapters.batch import (
+            substitute_hf_flash_attn,
+        )
 
-    substitute_hf_flash_attn(
-        process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride
-    )
+        substitute_hf_flash_attn(
+            process_group=get_ring_attn_group(),
+            ring_attn_func=ring_attn_func,
+        )
 
 
-def update_ring_attn_params(batch: dict[str, torch.Tensor]):
+def update_ring_attn_params(position_ids: torch.Tensor | None):
     """
     Calculate the cumulative sequence lengths for the current forward pass and pass the
     value to the substituted `ring_flash_attn`.
 
     Args:
-        batch: A dictionary with a batch of data. May or may not contain `position_ids`
-            data; if not, we compute it.
+        position_ids: Optional tensor of position IDs (for sample packed data).
     """
     from ring_flash_attn import update_ring_flash_attn_params
 
-    input_ids = batch["input_ids"]
-    position_ids = batch.get("position_ids")
-    if position_ids is None:
-        seq_len = input_ids.shape[1]
-        position_ids = torch.arange(
-            0, seq_len, dtype=torch.long, device=input_ids.device
-        ).unsqueeze(0)
-
     cu_seqlens, _ = get_cu_seqlens_from_pos_ids(position_ids)
     cu_seqlens = cu_seqlens.squeeze().to(device=torch.cuda.current_device())
     update_ring_flash_attn_params(cu_seqlens, get_ring_attn_group())

src/axolotl/monkeypatch/models/llama4/modeling.py

@@ -93,9 +93,20 @@ def patch_llama4_linearized_modeling():
     """
     from transformers.models.llama4 import modeling_llama4
 
+    old_lamma_4_text_experts = modeling_llama4.Llama4TextExperts
     modeling_llama4.Llama4TextExperts = Llama4TextExperts
     setattr(
         sys.modules["transformers.models.llama4"],
         "Llama4TextExperts",
         Llama4TextExperts,
     )
+
+    def unpatch():
+        modeling_llama4.Llama4TextExperts = old_lamma_4_text_experts
+        setattr(
+            sys.modules["transformers.models.llama4"],
+            "Llama4TextExperts",
+            old_lamma_4_text_experts,
+        )
+
+    return unpatch

src/axolotl/monkeypatch/multipack.py

@@ -31,6 +31,8 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "starcoder2",
     "deepseek_v2",
     "deepseek_v3",
+    "glm",
+    "glm4",
 ]
 
 

src/axolotl/monkeypatch/relora.py

@@ -272,7 +272,7 @@ class ReLoRAScheduler(LRScheduler):
         self.warmup_steps = warmup_steps
         self.anneal_steps = anneal_steps
         self.min_lr_scale = min_lr_scale
-        super().__init__(optimizer, inner_schedule.last_epoch, inner_schedule.verbose)
+        super().__init__(optimizer, inner_schedule.last_epoch)
 
     def get_lr(self) -> float:
         self.inner_schedule.last_epoch = self.last_epoch

src/axolotl/monkeypatch/trainer_eval_guard.py

@@ -0,0 +1,78 @@
+"""
+fix for FSDP2 evals when using torch.compile
+"""
+
+import inspect
+import logging
+
+from transformers import Trainer
+
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_TRAINER_CODE = """
+    model.eval()
+"""
+
+PATCHED_TRAINER_CODE = """
+    if hasattr(model, "eval") and callable(model.eval):
+        self.model.eval()
+"""
+
+
+def get_evaluation_loop_code() -> str:
+    training_loop = inspect.getsource(Trainer.evaluation_loop)
+    return training_loop
+
+
+def check_evaluation_loop_is_patchable() -> bool:
+    eval_loop = get_evaluation_loop_code()
+    eval_loop, _ = detab_code(eval_loop)
+    return ORIGINAL_TRAINER_CODE in eval_loop
+
+
+def patch_evaluation_loop_for_fsdp2():
+    """
+    monkeypatch for fixing the eval loop for fsdp2 with torch.compile
+    """
+
+    try:
+        evaluation_loop = get_evaluation_loop_code()
+    except OSError:
+        return
+    Trainer._original_evaluation_loop = (  # pylint: disable=protected-access
+        evaluation_loop
+    )
+    evaluation_loop, _ = detab_code(evaluation_loop)
+    if ORIGINAL_TRAINER_CODE not in evaluation_loop:
+        return
+
+    evaluation_loop = evaluation_loop.replace(
+        ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE
+    )
+    evaluation_loop = evaluation_loop.replace(
+        "def evaluation_loop(",
+        "def _fixed_evaluation_loop(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in evaluation_loop:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(evaluation_loop, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching _inner_training_loop for fsdp optimizer save")
+    Trainer.evaluation_loop = (  # pylint: disable=protected-access
+        _fixed_evaluation_loop  # pylint: disable=undefined-variable  # noqa: F821
+    )

src/axolotl/train.py

@@ -81,6 +81,11 @@ def setup_model_and_tokenizer(
     # Apply freezing if specified
     if cfg.unfrozen_parameters:
         freeze_layers_except(model, cfg.unfrozen_parameters)
+        if any(
+            any(embed in param for embed in ["lm_head", "embed_tokens"])
+            for param in cfg.unfrozen_parameters
+        ):
+            model.enable_input_require_grads()
 
     return model, tokenizer, peft_config, processor
 

src/axolotl/utils/collators/batching.py

@@ -4,7 +4,7 @@ includes logic for handling sequence parallelism collation.
 """
 
 from dataclasses import dataclass
-from typing import Any, Optional, Union
+from typing import Any
 
 import numpy as np
 import torch
@@ -13,6 +13,7 @@ from transformers import PreTrainedTokenizerBase
 from transformers.utils import PaddingStrategy
 
 from axolotl.monkeypatch.attention.ring_attn import update_ring_attn_params
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 
 @dataclass
@@ -53,14 +54,15 @@ class DataCollatorForSeq2Seq:
     """
 
     tokenizer: PreTrainedTokenizerBase
-    model: Optional[Any] = None
-    padding: Union[bool, str, PaddingStrategy] = True
-    max_length: Optional[int] = None
-    pad_to_multiple_of: Optional[int] = None
+    model: Any | None = None
+    padding: bool | str | PaddingStrategy = True
+    max_length: int | None = None
+    pad_to_multiple_of: int | None = None
     label_pad_token_id: int = -100
     position_pad_token_id: int = 0
     return_tensors: str = "pt"
     sequence_parallel_degree: int = 1
+    ring_attn_func: RingAttnFunc | None = None
 
     def __post_init__(self):
         if self.sequence_parallel_degree > 1:
@@ -157,19 +159,41 @@ class DataCollatorForSeq2Seq:
             Sliced batch dictionary.
         """
         # Get local (start, end) for sequence parallelism slicing
-        total_seq_len = batch["input_ids"].shape[1]
-        slice_size = total_seq_len // self.local_world_size
-        start = self.local_rank * slice_size
-        end = start + slice_size
+        total_seq_len = batch["input_ids"].size(1)
 
-        # Update params for ring attention calculation
-        update_ring_attn_params(batch=batch)
+        # Update params for varlen ring attention calculation
+        if batch.get("position_ids") is not None:
+            update_ring_attn_params(position_ids=batch["position_ids"])
 
         # Slice batch for sequence parallel processing
-        keys_to_slice = ["input_ids", "attention_mask", "labels", "position_ids"]
-        for key in keys_to_slice:
-            if key in batch:
-                batch[key] = batch[key][:, start:end]
+        for key in batch:
+            if batch[key].size(1) == total_seq_len:
+                if self.ring_attn_func in [
+                    RingAttnFunc.VARLEN_LLAMA3,
+                    RingAttnFunc.BATCH_RING,
+                ]:
+                    batch[key] = (
+                        batch[key]
+                        .chunk(self.local_world_size, dim=1)[self.local_rank]
+                        .contiguous()
+                    )
+                elif self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+                    chunks = batch[key].chunk(2 * self.local_world_size, dim=1)
+
+                    # Take rank's chunk and opposing chunk for zigzag pattern
+                    selected_chunks = [
+                        chunks[self.local_rank],
+                        chunks[2 * self.local_world_size - self.local_rank - 1],
+                    ]
+                    batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
+                elif self.ring_attn_func is RingAttnFunc.BATCH_STRIPE:
+                    # TODO(djsaunde): This doesn't seem to work as expected
+                    # Split into striped data and stack
+                    tensor = torch.stack(
+                        batch[key].split(self.local_world_size, dim=1),
+                        dim=1,
+                    ).transpose(1, 2)
+                    batch[key] = tensor[:, self.local_rank].contiguous()
 
         return batch
 

src/axolotl/utils/data/sft.py

@@ -3,6 +3,7 @@
 import functools
 import logging
 import os
+import tempfile
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
@@ -117,9 +118,27 @@ def prepare_dataset(cfg, tokenizer, processor=None, preprocess_iterable=None):
             cfg.pretraining_dataset[0]["type"] or "pretrain",
         )
 
-        iter_ds = load_dataset(
-            path, streaming=True, split=split, name=name, data_files=data_files
-        )
+        # when letting accelerator dispatch batches from the main process, we don't need to load the dataset from
+        # other ranks, we just need to present a fake dataset
+        if (
+            cfg.accelerator_config
+            and cfg.accelerator_config.dispatch_batches
+            and not is_local_main_process()
+        ):
+            with tempfile.NamedTemporaryFile(mode="w+", delete=False) as f:
+                f.write("text\n")
+                f.write("lorem ipsum dolor sit amet\n")
+                # rewind the file pointer to the beginning so we can read it again
+                f.seek(0)
+                iter_ds = load_dataset(
+                    "csv", data_files=f.name, split="train", streaming=True
+                )
+        else:
+            if is_local_main_process():
+                iter_ds = load_dataset(
+                    path, streaming=True, split=split, name=name, data_files=data_files
+                )
+
         if skip:
             LOG.info(f"Skipping {skip} samples from the dataset")
             iter_ds = iter_ds.skip(skip)
@@ -332,16 +351,23 @@ def load_tokenized_prepared_datasets(
         if cfg.local_rank == 0 and not cfg.skip_prepare_dataset:
             LOG.info(f"Saving merged prepared dataset to disk... {prepared_ds_path}")
             if isinstance(dataset, IterableDataset):
+                num_workers = cfg.dataset_processes
 
-                def gen_from_iter_ds(_ds, _=None):
-                    yield from _ds
+                def gen_from_iter_ds(_ds, worker_id: List[int], num_workers: List[int]):
+                    """Generator function to correctly splice the dataset for each worker"""
+                    for i, item in enumerate(_ds):
+                        if i % num_workers[0] == worker_id[0]:
+                            yield item
 
                 ds_from_iter = Dataset.from_generator(
                     functools.partial(gen_from_iter_ds, dataset),
                     features=dataset.features,
-                    num_proc=cfg.dataset_processes,
+                    num_proc=num_workers,
                     split=split,
-                    gen_kwargs={"_": list(range(cfg.dataset_processes))},
+                    gen_kwargs={
+                        "worker_id": list(range(num_workers)),
+                        "num_workers": [num_workers] * num_workers,
+                    },
                 )
                 ds_from_iter.save_to_disk(str(prepared_ds_path))
             else:

src/axolotl/utils/freeze.py

@@ -2,13 +2,14 @@
 module to freeze/unfreeze parameters by name
 """
 
-import logging
 import re
 from typing import Callable, List, Tuple, Union
 
+from accelerate.logging import get_logger
+
 from axolotl.utils.distributed import is_main_process
 
-LOG = logging.getLogger("axolotl.utils.freeze")
+LOG = get_logger(__name__)
 
 
 def freeze_layers_except(model, regex_patterns):
@@ -184,7 +185,7 @@ class LayerNamePattern:
         """
         self.raw_pattern = pattern
         name_pattern, self.range = self._parse_pattern(pattern)
-        self.name_regex = re.compile(name_pattern.replace(".", "\\."))
+        self.name_regex = re.compile(re.sub(r"\.(?!\+)", "\\.", name_pattern))
 
     def match(self, name: str) -> bool:
         """

src/axolotl/utils/models.py

@@ -542,6 +542,17 @@ class ModelLoader:
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 
             patch_accelerate_fsdp_utils()
+
+        if self.cfg.flex_attention:
+            from axolotl.monkeypatch.attention.flex_attn import (
+                patch_flex_make_mask,
+                patch_flex_wrapper,
+            )
+
+            flex_attn_compile_kwargs = self.cfg.flex_attn_compile_kwargs or {}
+            patch_flex_wrapper(**flex_attn_compile_kwargs)
+            patch_flex_make_mask()
+
         # patch gemma3 conditional generation forward before loading plugins
         # as it could be overridden by plugins
         if self.cfg.model_config_type == "llama4":
@@ -644,6 +655,7 @@ class ModelLoader:
             register_ring_attn(
                 sequence_parallel_degree=self.cfg.sequence_parallel_degree,
                 heads_k_stride=self.cfg.heads_k_stride,
+                ring_attn_func=self.cfg.ring_attn_func,
             )
 
     def patch_attention(self) -> None:
@@ -905,13 +917,6 @@ class ModelLoader:
             self.model_config._attn_implementation = (  # pylint: disable=protected-access
                 "flex_attention"
             )
-            from axolotl.monkeypatch.attention.flex_attn import (
-                patch_flex_make_mask,
-                patch_flex_wrapper,
-            )
-
-            patch_flex_wrapper()
-            patch_flex_make_mask()
 
         elif self.cfg.flash_attention:
             if not self.cfg.sample_packing and self.cfg.s2_attention:
@@ -1115,7 +1120,7 @@ class ModelLoader:
 
         return skip_move_to_device
 
-    def ajust_model_config(self) -> None:
+    def adjust_model_config(self) -> None:
         if (
             hasattr(self.model, "config")
             and hasattr(self.model.config, "max_position_embeddings")
@@ -1275,7 +1280,7 @@ class ModelLoader:
         else:
             self.model.tie_weights()
 
-        self.ajust_model_config()
+        self.adjust_model_config()
 
         # log device memory usage
         if hasattr(self.model, "device") and self.model.device.type in (

src/axolotl/utils/schedulers.py

@@ -40,7 +40,7 @@ class RexLR(LRScheduler):
         self.max_lr = max_lr
         self.total_steps = total_steps
         self.num_warmup_steps = num_warmup_steps
-        self.last_step = last_step - 1
+        self.last_step = max(last_step - 1, 0)
 
         # Ensure each parameter group has an "initial_lr" key to avoid issues when resuming.
         for group in optimizer.param_groups:

src/axolotl/utils/schemas/config.py

@@ -225,6 +225,7 @@ class AxolotlInputConfig(
     sdp_attention: bool | None = None
     s2_attention: bool | None = None
     flex_attention: bool | None = None
+    flex_attn_compile_kwargs: dict[str, Any] | None = None
     flash_attention: bool | None = None
     flash_attn_cross_entropy: bool | None = None
     flash_attn_rms_norm: bool | None = None
@@ -258,6 +259,7 @@ class AxolotlInputConfig(
 
     sequence_parallel_degree: int | None = None
     heads_k_stride: int | None = None
+    ring_attn_func: str | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -658,6 +660,7 @@ class AxolotlInputConfig(
             data.get("val_set_size") == 0
             and (data.get("eval_steps") or data.get("eval_strategy"))
             and not data.get("test_datasets")
+            and data.get("eval_strategy") != "no"
         ):
             raise ValueError(
                 "eval_steps and eval_strategy are not supported with val_set_size == 0"
@@ -1146,7 +1149,7 @@ class AxolotlInputConfig(
 
         return data
 
-    @field_validator("sequence_parallel_degree", mode="before")
+    @field_validator("sequence_parallel_degree", mode="after")
     @classmethod
     def check_sequence_parallel_degree(cls, value, info):
         if not value:
@@ -1158,9 +1161,12 @@ class AxolotlInputConfig(
                     "flash_attention: true must be set with sequence_parallel_degree > 1"
                 )
 
-            if not info.data["micro_batch_size"] == 1:
+            if (
+                info.data.get("sample_packing")
+                and not info.data["micro_batch_size"] == 1
+            ):
                 raise ValueError(
-                    "micro_batch_size must be set to 1 "
+                    "micro_batch_size must be set to 1 when sample_packing is enabled"
                     "due to a `ring-flash-attn` requirement"
                 )
 
@@ -1187,6 +1193,34 @@ class AxolotlInputConfig(
 
         return value
 
+    @field_validator("ring_attn_func", mode="after")
+    @classmethod
+    def check_ring_attn_func(cls, value, info):
+        if not info.data.get("sequence_parallel_degree", 1) > 1:
+            return value
+
+        from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
+        if value is not None:
+            # Set the ring attention function if passed in config
+            valid_funcs = list(RingAttnFunc)
+            if value in valid_funcs:
+                value = RingAttnFunc(value)
+            else:
+                raise ValueError(
+                    f"ring_attn_func: {value} must be one of {valid_funcs}"
+                )
+        else:
+            # Default ring attention function selection
+            sample_packing = info.data.get("sample_packing")
+            value = (
+                RingAttnFunc.VARLEN_LLAMA3
+                if sample_packing
+                else RingAttnFunc.BATCH_RING
+            )
+
+        return value
+
     @model_validator(mode="before")
     @classmethod
     def check_muon_deepspeed_fsdp(cls, data):
@@ -1276,11 +1310,14 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
         ):
             capabilities = data.get("capabilities")
             is_fsdp = data.get("fsdp") is not None
-
-            if capabilities and capabilities.get("n_gpu", 0) > 1:
+            is_fsdp2 = (
+                data.get("fsdp_config") is not None
+                and str(data.get("fsdp_config").get("fsdp_version")) == "2"
+            )
+            if capabilities and capabilities.get("n_gpu", 0) > 1 and not is_fsdp2:
                 if is_fsdp:
                     raise ValueError(
-                        "lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with FSDP."
+                        "lora_mlp_kernel, lora_qkv_kernel, and lora_o_kernel are not compatible with FSDP1."
                     )
         return data
 

src/axolotl/utils/schemas/vllm.py

@@ -36,3 +36,11 @@ class VllmConfig(BaseModel):
         default=None,
         json_schema_extra={"description": "Enable prefix caching for VLLM"},
     )
+    host: str | None = Field(
+        default="0.0.0.0",  # nosec B104
+        json_schema_extra={"description": "Host for the vLLM server to start on"},
+    )
+    port: int | None = Field(
+        default=8000,
+        json_schema_extra={"description": "Port of the vLLM server to start on"},
+    )

src/axolotl/utils/trainer.py

@@ -17,6 +17,7 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.monkeypatch.trainer_eval_guard import patch_evaluation_loop_for_fsdp2
 from axolotl.utils.distributed import reduce_and_broadcast
 from axolotl.utils.environment import check_cuda_p2p_ib_support
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
@@ -235,7 +236,8 @@ def drop_long_seq(sample, sequence_len=2048, min_sequence_len=2):
 
 
 def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
-    if cfg.model_config_type in ["mamba", "gemma3"]:
+    drop_attn_mask = cfg.model_config_type in ["mamba", "gemma3"]
+    if drop_attn_mask:
         LOG.info("dropping attention_mask column")
         train_dataset = train_dataset.remove_columns("attention_mask")
         if eval_dataset:
@@ -625,6 +627,12 @@ def setup_trainer(
         A trainer instance (either `HFRLTrainer` or `HFCausalTrainer`) configured based
             on the provided parameters.
     """
+    if (
+        cfg.torch_compile
+        and cfg.fsdp_config
+        and str(cfg.fsdp_config.fsdp_version) == "2"
+    ):
+        patch_evaluation_loop_for_fsdp2()
     if cfg.rl:
         trainer_builder = HFRLTrainerBuilder(cfg, model, tokenizer, processor)
         trainer_builder.model_ref = model_ref

tests/conftest.py

@@ -193,6 +193,14 @@ def download_tiny_shakespeare_dataset():
     snapshot_download_w_retry("winglian/tiny-shakespeare", repo_type="dataset")
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_evolkit_kd_sample_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample", repo_type="dataset"
+    )
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_deepseek_model_fixture():
     snapshot_download_w_retry("axolotl-ai-co/DeepSeek-V3-11M", repo_type="model")
@@ -208,6 +216,16 @@ def download_huggyllama_model_fixture():
     )
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_llama33_70b_model_fixture():
+    # download the tokenizer only
+    snapshot_download_w_retry(
+        "axolotl-ai-co/Llama-3.3-70B-Instruct-tokenizer",
+        repo_type="model",
+        allow_patterns=["*token*", "config.json"],
+    )
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_llama_1b_model_fixture():
     # download the tokenizer only
@@ -315,6 +333,14 @@ def download_llama2_model_fixture():
     )
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_llama32_1b_model_fixture():
+    snapshot_download_w_retry(
+        "osllmai-community/Llama-3.2-1B",
+        repo_type="model",
+    )
+
+
 @pytest.fixture
 @enable_hf_offline
 def tokenizer_huggyllama(

tests/e2e/integrations/test_cut_cross_entropy.py

@@ -8,7 +8,7 @@ from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils import get_pytorch_version
-from axolotl.utils.config import normalize_config, prepare_plugins
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists
@@ -56,6 +56,7 @@ class TestCutCrossEntropyIntegration:
     # pylint: disable=redefined-outer-name
     def test_llama_w_cce(self, min_cfg, temp_dir):
         cfg = DictDefault(min_cfg)
+        cfg = validate_config(cfg)
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
@@ -101,6 +102,7 @@ class TestCutCrossEntropyIntegration:
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
@@ -129,6 +131,7 @@ class TestCutCrossEntropyIntegration:
                 attention_type: True,
             }
         )
+        cfg = validate_config(cfg)
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()

tests/e2e/integrations/test_liger.py

@@ -5,7 +5,7 @@ Simple end-to-end test for Liger integration
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config, prepare_plugins
+from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
 from tests.e2e.utils import check_model_output_exists, require_torch_2_4_1
@@ -54,6 +54,7 @@ class LigerIntegrationTestCase:
             }
         )
         # pylint: disable=duplicate-code
+        cfg = validate_config(cfg)
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
@@ -100,6 +101,7 @@ class LigerIntegrationTestCase:
             }
         )
         # pylint: disable=duplicate-code
+        cfg = validate_config(cfg)
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()

tests/e2e/multigpu/patched/test_sp.py

@@ -3,13 +3,14 @@
 import os
 from pathlib import Path
 
+import pytest
 import yaml
 from accelerate.test_utils import execute_subprocess_async
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_tensorboard
+from ...utils import check_tensorboard
 
 os.environ["WANDB_DISABLED"] = "true"
 
@@ -17,8 +18,15 @@ os.environ["WANDB_DISABLED"] = "true"
 class TestSequenceParallelism:
     """Test case for training with sequence parallelism enabled"""
 
-    def test_sequence_parallel_training(self, temp_dir):
-        # pylint: disable=duplicate-code
+    def _run_sequence_parallel_test(
+        self,
+        temp_dir,
+        sample_packing=True,
+        micro_batch_size=1,
+        pad_to_sequence_len=True,
+        ring_attn_func=None,
+    ):
+        """Helper method to run sequence parallel tests with different configurations"""
         cfg = DictDefault(
             {
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
@@ -27,9 +35,9 @@ class TestSequenceParallelism:
                 "strict": False,
                 "sequence_len": 2048,
                 "adapter": "qlora",
-                "sample_packing": True,
-                "eval_sample_packing": True,
-                "pad_to_sequence_len": True,
+                "sample_packing": sample_packing,
+                "eval_sample_packing": sample_packing,
+                "pad_to_sequence_len": pad_to_sequence_len,
                 "lora_r": 8,
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
@@ -45,7 +53,7 @@ class TestSequenceParallelism:
                 ],
                 "num_epochs": 1,
                 "max_steps": 8,
-                "micro_batch_size": 1,
+                "micro_batch_size": micro_batch_size,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
@@ -61,6 +69,7 @@ class TestSequenceParallelism:
                 "weight_decay": 0.0,
                 "use_tensorboard": True,
                 "sequence_parallel_degree": 2,
+                "ring_attn_func": ring_attn_func,
             }
         )
 
@@ -86,3 +95,35 @@ class TestSequenceParallelism:
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
         )
+
+    @pytest.mark.parametrize(
+        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func",
+        [
+            (True, 1, True, None),  # defaults to varlen_llama3 ring_attn_func
+            (False, 2, True, None),  # defaults to batch_ring ring_attn_func
+            (False, 2, True, "batch_zigzag"),
+            # (False, 2, False),  # not yet working
+        ],
+        ids=[
+            "sample_packing, varlen_llama3 ring_attn_func",
+            "no sample_packing, no pad_to_sequence_len, batch_ring ring_attn_func",
+            "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
+            # "no sample_packing, pad_to_sequence_len",  # not yet working
+        ],
+    )
+    def test_sequence_parallel_training(
+        self,
+        temp_dir,
+        sample_packing,
+        micro_batch_size,
+        pad_to_sequence_len,
+        ring_attn_func,
+    ):
+        """Test sequence parallel training with different configurations"""
+        self._run_sequence_parallel_test(
+            temp_dir,
+            sample_packing=sample_packing,
+            micro_batch_size=micro_batch_size,
+            pad_to_sequence_len=pad_to_sequence_len,
+            ring_attn_func=ring_attn_func,
+        )

tests/e2e/multigpu/solo/__init__.py

@@ -0,0 +1,2 @@
+# Tests under this directory should get run "solo" on their own as they
+# seem to cause issues when run in the same batch as other tests.

tests/e2e/multigpu/solo/test_flex.py

@@ -49,18 +49,20 @@ class TestPackedFlex:
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 2,
                 "use_tensorboard": True,
                 "save_strategy": "no",
             }

tests/e2e/multigpu/solo/test_grpo.py

@@ -177,6 +177,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
             "NCCL_P2P_LEVEL": "LOC",
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
+            "VLLM_USE_V1": "0",
         }
         vllm_process_id = start_vllm(
             cfg.base_model,
@@ -264,6 +265,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
             "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
+            "VLLM_USE_V1": "0",
         }
         vllm_process_id = start_vllm(
             cfg.base_model,

tests/e2e/multigpu/test_llama.py

@@ -621,12 +621,6 @@ class TestMultiGPULlama:
             temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
         )
 
-    # TODO: remove skip once deepspeed regression is fixed
-    # see https://github.com/huggingface/transformers/pull/37324
-    @pytest.mark.skipif(
-        transformers_version_eq("4.51.0"),
-        reason="zero3 is not supported with transformers==4.51.0",
-    )
     @pytest.mark.parametrize(
         "gradient_accumulation_steps",
         [1, 2],

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -144,7 +144,7 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="auto"
+        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda:0"
     )
     peft_config = get_peft_config(
         {
@@ -347,7 +347,7 @@ def test_model_architecture(model_config):
     """Test LoRA kernel patches across different model architectures."""
     # Load model with appropriate dtype
     model = AutoModelForCausalLM.from_pretrained(
-        model_config["name"], torch_dtype=model_config["dtype"], device_map="auto"
+        model_config["name"], torch_dtype=model_config["dtype"], device_map="cuda:0"
     )
 
     # Apply LoRA configuration

tests/e2e/patched/test_4d_multipack_llama.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -60,6 +60,7 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "fp16": True,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -104,6 +105,7 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "fp16": True,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_falcon_samplepack.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -63,6 +63,7 @@ class TestFalconPatched(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -103,6 +104,7 @@ class TestFalconPatched(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_fused_llama.py

@@ -12,7 +12,7 @@ from transformers.utils import is_torch_bf16_gpu_available
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -67,6 +67,7 @@ class TestFusedLlama(unittest.TestCase):
             cfg.bf16 = True
         else:
             cfg.fp16 = True
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_llama_s2_attention.py

@@ -11,7 +11,7 @@ import pytest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -65,6 +65,7 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -105,6 +106,7 @@ class TestLlamaShiftedSparseAttention(unittest.TestCase):
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_lora_llama_multipack.py

@@ -12,7 +12,7 @@ from transformers.utils import is_auto_gptq_available, is_torch_bf16_gpu_availab
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -70,6 +70,7 @@ class TestLoraLlama(unittest.TestCase):
         else:
             cfg.fp16 = True
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -120,6 +121,7 @@ class TestLoraLlama(unittest.TestCase):
                 "lr_scheduler": "cosine",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_mistral_samplepack.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -63,6 +63,7 @@ class TestMistral(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -104,6 +105,7 @@ class TestMistral(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_mixtral_samplepack.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -60,6 +60,7 @@ class TestMixtral(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_model_patches.py

@@ -6,7 +6,7 @@ import unittest
 
 import transformers
 
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
 
@@ -47,6 +47,7 @@ class TestModelPatches(unittest.TestCase):
                 "eval_steps": 10,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         tokenizer = load_tokenizer(cfg)
         load_model(cfg, tokenizer, inference=False)
@@ -79,6 +80,7 @@ class TestModelPatches(unittest.TestCase):
                 "eval_steps": 10,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         tokenizer = load_tokenizer(cfg)
         load_model(cfg, tokenizer, inference=False)

tests/e2e/patched/test_phi_multipack.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, with_temp_dir
@@ -63,6 +63,7 @@ class TestPhiMultipack(unittest.TestCase):
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -82,7 +83,7 @@ class TestPhiMultipack(unittest.TestCase):
                 "sample_packing": True,
                 "flash_attention": True,
                 "pad_to_sequence_len": True,
-                "load_in_8bit": False,
+                "load_in_4bit": True,
                 "adapter": "qlora",
                 "lora_r": 64,
                 "lora_alpha": 32,
@@ -114,6 +115,7 @@ class TestPhiMultipack(unittest.TestCase):
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_resume.py

@@ -12,7 +12,7 @@ from transformers.utils import is_torch_bf16_gpu_available
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, most_recent_subdir
@@ -46,8 +46,9 @@ class TestResumeLlama:
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 2,
@@ -67,6 +68,7 @@ class TestResumeLlama:
             cfg.bf16 = True
         else:
             cfg.fp16 = True
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/patched/test_sp.py

@@ -73,7 +73,10 @@ class TestRingAttention:
         self, mock_world_size, mock_rank, mock_new_group, partial_state
     ):
         """Test that ring attention groups are created correctly."""
-        from axolotl.monkeypatch.attention.ring_attn import register_ring_attn
+        from axolotl.monkeypatch.attention.ring_attn import (
+            RingAttnFunc,
+            register_ring_attn,
+        )
 
         # Setup mocks
         mock_world_size.return_value = 8  # 8 GPUs total
@@ -82,7 +85,11 @@ class TestRingAttention:
         mock_new_group.return_value = mock_group
 
         # Call register_ring_attn with size 4
-        register_ring_attn(sequence_parallel_degree=4, heads_k_stride=1)
+        register_ring_attn(
+            sequence_parallel_degree=4,
+            heads_k_stride=1,
+            ring_attn_func=RingAttnFunc.VARLEN_LLAMA3,
+        )
 
         # Verify the number of calls without examining the arguments
         assert mock_new_group.call_count == 2

tests/e2e/patched/test_unsloth_qlora.py

@@ -10,7 +10,7 @@ import pytest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from ..utils import check_model_output_exists, check_tensorboard
@@ -72,6 +72,7 @@ class TestUnslothQLoRA:
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -122,6 +123,7 @@ class TestUnslothQLoRA:
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -177,6 +179,7 @@ class TestUnslothQLoRA:
             }
         )
 
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/solo/test_flex.py

@@ -41,8 +41,9 @@ class TestPackedFlex(unittest.TestCase):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,

tests/e2e/test_embeddings_lr.py

@@ -102,6 +102,7 @@ class TestEmbeddingsLrScale(unittest.TestCase):
                 "use_tensorboard": True,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/test_llama_vision.py

@@ -109,6 +109,7 @@ class TestLlamaVision(unittest.TestCase):
                 "bf16": True,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/test_packing_loss.py

@@ -40,8 +40,9 @@ class TestPackedLlama(unittest.TestCase):
                 },
                 "datasets": [
                     {
-                        "path": "vicgalle/alpaca-gpt4",
+                        "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:10%]",
                     },
                 ],
                 "num_epochs": 1,

tests/e2e/test_phi.py

@@ -79,7 +79,7 @@ class TestPhi(unittest.TestCase):
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 2048,
                 "sample_packing": False,
-                "load_in_8bit": False,
+                "load_in_4bit": True,
                 "adapter": "qlora",
                 "lora_r": 64,
                 "lora_alpha": 32,
@@ -111,6 +111,7 @@ class TestPhi(unittest.TestCase):
                 "bf16": "auto",
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/test_process_reward_model_smollm2.py

@@ -9,7 +9,7 @@ import unittest
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
 from .utils import check_model_output_exists, check_tensorboard, with_temp_dir
@@ -57,6 +57,7 @@ class TestProcessRewardSmolLM2(unittest.TestCase):
                 "seed": 42,
             }
         )
+        cfg = validate_config(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/test_exact_deduplication.py

@@ -11,7 +11,7 @@ from unittest.mock import patch
 import pytest
 from datasets import Dataset
 
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.data.utils import deduplicate_and_log_datasets
@@ -319,6 +319,7 @@ class TestDeduplicateNonRL(unittest.TestCase):
                 "num_epochs": 1,
             }
         )
+        self.cfg_1 = validate_config(self.cfg_1)
         normalize_config(self.cfg_1)
 
     @pytest.mark.skip(reason="TODO: fix hf hub offline to work with HF rate limits")