Merge branch 'main' into codecov-pulls-only

.github/workflows/tests.yml

@@ -106,13 +106,12 @@ jobs:
           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
+      - name: Upload coverage artifacts
-        uses: codecov/codecov-action@v5
+        uses: actions/upload-artifact@v4
         with:
-          token: ${{ secrets.CODECOV_TOKEN }}
+          name: coverage-${{ matrix.pytorch_version }}-${{ github.run_id }}
-          files: ./coverage.xml
+          path: ./coverage.xml
-          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          retention-days: 1
-          fail_ci_if_error: false
 
       - name: cleanup pip cache
         run: |
@@ -234,6 +233,14 @@ jobs:
         run: |
           modal run cicd.e2e_tests
 
+      - name: Upload coverage artifacts
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: coverage-e2e-1st-${{ github.run_id }}
+          path: ./e2e-coverage.xml
+          retention-days: 1
+
   docker-e2e-tests:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
@@ -297,6 +304,14 @@ jobs:
         run: |
           modal run cicd.e2e_tests
 
+      - name: Upload coverage artifacts
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: coverage-e2e-${{ matrix.cuda }}-${{ matrix.pytorch }}-${{ github.run_id }}
+          path: ./e2e-coverage.xml
+          retention-days: 1
+
   docker-e2e-cleanup:
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
@@ -336,3 +351,26 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.cleanup
+
+  upload-coverage:
+    name: Upload Coverage to Codecov
+    runs-on: ubuntu-latest
+    needs: [pytest, docker-e2e-tests, docker-e2e-tests-1st]
+    if: github.event_name == 'pull_request' || github.ref == 'refs/heads/main'
+
+    steps:
+      - name: Download coverage reports
+        uses: actions/download-artifact@v4
+        with:
+          path: coverage-reports
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          directory: coverage-reports
+          fail_ci_if_error: false
+          verbose: true
+          name: codecov-umbrella
+          override_commit: ${{ github.event.pull_request.head.sha || github.sha }}
+          override_pr: ${{ github.event.pull_request.number }}

.pre-commit-config.yaml

@@ -19,7 +19,7 @@ repos:
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 7.3.0
+    rev: 7.2.0
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
@@ -27,7 +27,7 @@ repos:
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.1
+    rev: v1.16.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.5
+    rev: 1.8.3
     hooks:
     -   id: bandit
         args: [

README.md

@@ -43,7 +43,7 @@ Features:
 - **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
 - **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), Sequence Parallelism (SP), LoRA optimizations, Multi-GPU training (FSDP1, FSDP2, DeepSpeed), Multi-node training (Torchrun, Ray), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 

cicd/cicd.sh

@@ -51,5 +51,3 @@ pytest -v --durations=10 \
   --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} || true

cicd/e2e_tests.py

@@ -1,5 +1,7 @@
 """Modal app to run axolotl GPU tests"""
 
+import pathlib
+
 from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 
 
@@ -12,9 +14,21 @@ from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
     volumes=VOLUME_CONFIG,
 )
 def cicd_pytest():
+
     run_cmd("./cicd/cicd.sh", "/workspace/axolotl")
 
+    # Read the coverage file if it exists
+    coverage_file = pathlib.Path("/workspace/axolotl/e2e-coverage.xml")
+    if coverage_file.exists():
+        return coverage_file.read_text(encoding="utf-8")
+    return None
+
 
 @app.local_entrypoint()
 def main():
-    cicd_pytest.remote()
+    coverage = cicd_pytest.remote()
+
+    # Save the coverage file to the local filesystem if it was generated
+    if coverage:
+        with open("e2e-coverage.xml", "w", encoding="utf-8") as f:
+            f.write(coverage)

cicd/multigpu.py

@@ -77,7 +77,18 @@ def run_cmd(cmd: str, run_folder: str):
 def cicd_pytest():
     run_cmd("./cicd/multigpu.sh", "/workspace/axolotl")
 
+    # Read the coverage file if it exists
+    coverage_file = pathlib.Path("/workspace/axolotl/multigpu-coverage.xml")
+    if coverage_file.exists():
+        return coverage_file.read_text(encoding="utf-8")
+    return None
+
 
 @app.local_entrypoint()
 def main():
-    cicd_pytest.remote()
+    coverage = cicd_pytest.remote()
+
+    # Save the coverage file to the local filesystem if it was generated
+    if coverage:
+        with open("multigpu-coverage.xml", "w", encoding="utf-8") as file:
+            file.write(coverage)

docs/dataset-formats/conversation.qmd

@@ -9,7 +9,7 @@ order: 3
 Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
 
 ```{.json filename="data.jsonl"}
-{"messages": [{"role": "...", "content": "..."}, {"role": "...", "content": "..."}, ...]}
+{"conversations": [{"role": "...", "content": "..."}]}
 ```
 
 See [configs](../config-reference.qmd) for full configs and supported templates.

docs/faq.qmd

@@ -9,11 +9,11 @@ description: Frequently asked questions
 
 > A: Usually an issue with the GPUs communicating with each other. See the [NCCL doc](nccl.qmd)
 
-**Q: exitcode: -9**
+**Q: Exitcode -9**
 
 > A: This usually happens when you run out of system RAM.
 
-**Q: exitcode: -7 while using deepspeed**
+**Q: Exitcode -7 while using deepspeed**
 
 > A: Try upgrading deepspeed w: `pip install -U deepspeed`
 

examples/falcon-h1/falcon-h1-1b-deep-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-1.5B-Deep-Base
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/falcon-h1/falcon-h1-1b-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-1.5B-Base
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/falcon-h1/falcon-h1-34b-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-34B-Base
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/falcon-h1/falcon-h1-3b-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-3B-Base
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/falcon-h1/falcon-h1-500m-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-0.5B-Instruct
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/falcon-h1/falcon-h1-7b-qlora.yaml

@@ -1,71 +0,0 @@
-base_model: tiiuae/Falcon-H1-7B-Base
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: true
-
-# huggingface repo
-chat_template: falcon_h1
-datasets:
-  - path: cgato/SlimOrcaDedupCleaned
-    type: chat_template
-    field_messages: conversations
-    message_property_mappings:
-      role: from
-      content: value
-
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - k_proj
-  - v_proj
-  - o_proj
-  - in_proj
-  - gate_proj
-  - up_proj
-  - down_proj
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:

examples/gemma3/gemma-3-1b-qlora.yml

@@ -13,8 +13,6 @@ load_in_4bit: true
 
 # huggingface repo
 chat_template: gemma3
-eot_tokens:
-  - <end_of_turn>
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template

examples/gemma3/gemma-3-4b-qlora.yml

@@ -6,8 +6,6 @@ load_in_4bit: true
 ddp_find_unused_parameters: true
 
 chat_template: gemma3
-eot_tokens:
-  - <end_of_turn>
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -12,8 +12,6 @@ sample_packing: false
 ddp_find_unused_parameters: true
 
 chat_template: gemma3
-eot_tokens:
-  - <end_of_turn>
 datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template

examples/qwen2_5-vl/lora-7b.yaml

@@ -1,55 +0,0 @@
-base_model: Qwen/Qwen2.5-VL-7B-Instruct
-processor_type: AutoProcessor
-
-# 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
-
-chat_template: qwen2_vl
-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
-
-adapter: lora
-lora_model_dir:
-
-sequence_len: 8192
-pad_to_sequence_len: false
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: true
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-logging_steps: 1
-flash_attention: true
-eager_attention:
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0

requirements.txt

@@ -18,7 +18,7 @@ tokenizers>=0.21.1
 accelerate==1.7.0
 datasets==3.6.0
 deepspeed>=0.17.0
-trl==0.18.2
+trl==0.18.1
 hf_xet==1.1.2
 
 optimum==1.16.2

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@78b2a45713a54c9bedf8b33f5e31cf07a1a57154"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@a1174ca"'
 )

src/axolotl/cli/checks.py

@@ -6,7 +6,6 @@ from pathlib import Path
 from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
-from requests import HTTPError
 
 from axolotl.utils.logging import get_logger
 
@@ -47,8 +46,3 @@ def check_user_token() -> bool:
             "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
         )
         return False
-    except HTTPError:
-        LOG.warning(
-            "Error accessing HuggingFace. This may be due to a network issue or rate limiting."
-        )
-        return False

src/axolotl/cli/cloud/__init__.py

@@ -7,6 +7,7 @@ from typing import Union
 
 import yaml
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.cloud.modal_ import ModalCloud
 from axolotl.utils.dict import DictDefault
 
@@ -23,6 +24,7 @@ def do_cli_preprocess(
     cloud_config: Union[Path, str],
     config: Union[Path, str],
 ) -> None:
+    print_axolotl_text_art()
     cloud_cfg = load_cloud_cfg(cloud_config)
     cloud = ModalCloud(cloud_cfg)
     with open(config, "r", encoding="utf-8") as file:
@@ -37,6 +39,7 @@ def do_cli_train(
     cwd=None,
     **kwargs,
 ) -> None:
+    print_axolotl_text_art()
     cloud_cfg = load_cloud_cfg(cloud_config)
     cloud = ModalCloud(cloud_cfg)
     with open(config, "r", encoding="utf-8") as file:
@@ -51,6 +54,7 @@ def do_cli_lm_eval(
     cloud_config: Union[Path, str],
     config: Union[Path, str],
 ) -> None:
+    print_axolotl_text_art()
     cloud_cfg = load_cloud_cfg(cloud_config)
     cloud = ModalCloud(cloud_cfg)
     with open(config, "r", encoding="utf-8") as file:

src/axolotl/cli/config.py

@@ -26,9 +26,7 @@ from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.trainer import prepare_opinionated_env, prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
-
-API_KEY_FIELDS = {"comet_api_key"}
 
 
 def check_remote_config(config: Union[str, Path]) -> Union[str, Path]:
@@ -235,15 +233,4 @@ def load_cfg(
     setup_comet_env_vars(cfg)
     plugin_set_cfg(cfg)
 
-    cfg_to_log = {
-        k: "[REDACTED]" if k in API_KEY_FIELDS else v
-        for k, v in cfg.items()
-        if v is not None
-    }
-
-    LOG.info(
-        "config:\n%s",
-        json.dumps(cfg_to_log, indent=2, default=str, sort_keys=True),
-    )
-
     return cfg

src/axolotl/cli/evaluate.py

@@ -9,6 +9,7 @@ from dotenv import load_dotenv
 from transformers.hf_argparser import HfArgumentParser
 
 from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.datasets import load_datasets, load_preference_datasets
@@ -34,6 +35,7 @@ def do_evaluate(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
     patch_optimized_env()
 
     # pylint: disable=duplicate-code
+    print_axolotl_text_art()
     check_accelerate_default_config()
     if int(os.getenv("LOCAL_RANK", "0")) == 0:
         check_user_token()

src/axolotl/cli/inference.py

@@ -13,6 +13,7 @@ from dotenv import load_dotenv
 from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
 
 from axolotl.cli.args import InferenceCliArgs
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
 from axolotl.utils.chat_templates import (
@@ -254,6 +255,7 @@ def do_cli(
         kwargs: Additional keyword arguments to override config file values.
     """
     # pylint: disable=duplicate-code
+    print_axolotl_text_art()
     parsed_cfg = load_cfg(config, inference=True, rl=None, **kwargs)
     parsed_cfg.sample_packing = False
     parser = transformers.HfArgumentParser(InferenceCliArgs)

src/axolotl/cli/main.py

@@ -20,7 +20,6 @@ from axolotl.cli.args import (
     TrainerCliArgs,
     VllmServeCliArgs,
 )
-from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.sweeps import generate_sweep_configs
 from axolotl.cli.utils import (
     add_options_from_config,
@@ -41,7 +40,6 @@ LOG = get_logger(__name__)
 @click.version_option(version=axolotl.__version__, prog_name="axolotl")
 def cli():
     """Axolotl CLI - Train and fine-tune large language models"""
-    print_axolotl_text_art()
 
 
 @cli.command()

src/axolotl/cli/merge_lora.py

@@ -6,6 +6,7 @@ from typing import Union
 import fire
 from dotenv import load_dotenv
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
 from axolotl.utils.dict import DictDefault
@@ -22,6 +23,8 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
     """
+    print_axolotl_text_art()
+
     model, tokenizer, processor = load_model_and_tokenizer(cfg=cfg)
     safe_serialization = cfg.save_safetensors is True
 

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -22,6 +22,7 @@ from huggingface_hub import split_torch_state_dict_into_shards
 from safetensors.torch import save_file as safe_save_file
 from torch.distributed.checkpoint.format_utils import _EmptyStateDictLoadPlanner
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.utils.logging import get_logger
 
@@ -193,6 +194,7 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         kwargs: Additional keyword arguments to override config file values.
     """
     # pylint: disable=duplicate-code
+    print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
 
     fsdp_dir = Path(parsed_cfg.output_dir) / "pytorch_model_fsdp_0"

src/axolotl/cli/preprocess.py

@@ -12,6 +12,7 @@ from dotenv import load_dotenv
 from transformers import AutoModelForCausalLM
 
 from axolotl.cli.args import PreprocessCliArgs
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
@@ -32,6 +33,7 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: Preprocessing-specific CLI arguments.
     """
+    print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
 

src/axolotl/cli/quantize.py

@@ -7,6 +7,7 @@ from typing import Union
 
 from transformers import AutoModelForCausalLM
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.loaders import load_tokenizer
 from axolotl.utils.logging import get_logger
@@ -26,6 +27,7 @@ def do_quantize(
         config (Union[Path, str]): The path to the config file
         cli_args (dict): Additional command-line arguments
     """
+    print_axolotl_text_art()
 
     cfg = load_cfg(config)
 

src/axolotl/cli/train.py

@@ -11,6 +11,7 @@ from dotenv import load_dotenv
 from transformers.hf_argparser import HfArgumentParser
 
 from axolotl.cli.args import TrainerCliArgs
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.datasets import load_datasets, load_preference_datasets
@@ -34,6 +35,7 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
     # Enable expandable segments for cuda allocation to improve VRAM usage
     patch_optimized_env()
 
+    print_axolotl_text_art()
     check_accelerate_default_config()
     if int(os.getenv("LOCAL_RANK", "0")) == 0:
         check_user_token()

src/axolotl/core/trainers/base.py

@@ -20,7 +20,7 @@ from torch.utils.data import (
     SequentialSampler,
 )
 from transformers import Trainer
-from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length, seed_worker
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, seed_worker
 from trl.trainer.utils import pad_to_length
 from typing_extensions import override
 
@@ -116,15 +116,14 @@ class AxolotlTrainer(
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
             num_processes=self.args.dataset_num_proc,
-            mp_start_method=self.args.sample_packing_mp_start_method or "fork",
         )
 
         len(sampler)
         return sampler
 
     def _get_train_sampler(
-        self, train_dataset: Dataset | None = None
+        self, train_dataset: Optional[Dataset] = None
-    ) -> Sampler | None:
+    ) -> Optional[Sampler]:
         """
         Helper method to get the sampler for training. Handles cases for sample packing
         and curriculum sampling (sequential).
@@ -133,22 +132,16 @@ class AxolotlTrainer(
             If the dataset is non-empty, a sampler is returned, the type of which
                 depends on the passed training args.
         """
-        # from https://github.com/huggingface/transformers/blob/2166b6b4ff09f6dd3867ab982f262f66482aa968/src/transformers/trainer.py#L969C1-L972C24
-        if train_dataset is None:
-            train_dataset = self.train_dataset
-        if train_dataset is None or not has_length(train_dataset):
-            return None
-
         use_sample_packing = self.args.sample_packing and not self.args.pretraining
 
         # Determine the base sampler first
         if self.args.curriculum_sampling:
-            base_sampler = SequentialSampler(train_dataset)
+            base_sampler = SequentialSampler(self.train_dataset)
         elif use_sample_packing:
-            base_sampler = RandomSampler(train_dataset)
+            base_sampler = RandomSampler(self.train_dataset)
         else:
             # Default to parent class implementation for standard random sampling
-            return super()._get_train_sampler(train_dataset)
+            return super()._get_train_sampler()
 
         # Apply multipack wrapper if needed
         if use_sample_packing:
@@ -167,10 +160,6 @@ class AxolotlTrainer(
             If the dataset is non-empty, a sampler is returned, the type of which
                 depends on the passed training args.
         """
-        # from https://github.com/huggingface/transformers/blob/2166b6b4ff09f6dd3867ab982f262f66482aa968/src/transformers/trainer.py#L1065C9-L1066C24
-        if eval_dataset is None or not has_length(eval_dataset):
-            return None
-
         # Multipacking enabled if training is enabled and eval is not explicitly disabled
         use_multipack = (
             self.args.sample_packing and self.args.eval_sample_packing is not False

src/axolotl/core/training_args_base.py

@@ -38,10 +38,6 @@ class AxolotlTrainingMixins:
             "help": "Use next-fit sample packing that preserves the order of samples coming from the sampler. Use in combination with curriculum_sampling for fully sequential packing."
         },
     )
-    sample_packing_mp_start_method: str | None = field(
-        default=None,
-        metadata={"help": "The multiprocessing start method to use."},
-    )
     multipack_real_batches: bool = field(
         default=False,
         metadata={"help": "Use real batches for efficient training."},

src/axolotl/integrations/base.py

@@ -33,7 +33,7 @@ from transformers import PreTrainedModel, Trainer
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 if TYPE_CHECKING:
     from axolotl.common.datasets import TrainDatasetMeta

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,11 +19,19 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@78b2a45713a54c9bedf8b33f5e31cf07a1a57154"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"
 ```
 
 ## Usage
 
+**NOTE**: If you are training a VLM model, please use older version of Axolotl as upstream has applied a major VLM refactor, and our patches have not been updated yet.
+
+```bash
+git checkout 787880215b3ab32ccaf81c1b2e9588c6f3e6e764
+
+pip3 install --no-build-isolation -e .
+```
+
 ```yaml
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
@@ -31,29 +39,27 @@ plugins:
 
 ## Supported Models
 
-- cohere
+- llama
-- cohere2
+- llama4
+- llama4_text
+- mllama
+- phi3
 - gemma
 - gemma2
 - gemma3
 - gemma3_text
-- glm
-- glm4
-- llama
-- llama4
-- llama4_text
 - mistral
 - mistral3
-- mllama
-- phi
-- phi3
-- phi4_multimodal
 - qwen2
-- qwen2_vl
 - qwen2_moe
+- qwen2_vl
 - qwen2_5_vl
 - qwen3
 - qwen3_moe
+- cohere
+- cohere2
+- glm
+- glm4
 
 ## Citation
 

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -28,11 +28,11 @@ from axolotl.utils.logging import get_logger
 
 from .args import CutCrossEntropyArgs  # pylint: disable=unused-import. # noqa: F401
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 _CCE_INSTALL_MESSAGE = (
-    "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
+    "Please install cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@7f6afce"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"`'
 )
 
 
@@ -64,28 +64,16 @@ class CutCrossEntropyPlugin(BasePlugin):
             "cut_cross_entropy.transformers"
         )
         if cce_spec_transformers is None:
-            raise ImportError(
+            raise ImportError(_CCE_INSTALL_MESSAGE)
-                "Transformers support is not installed. " + _CCE_INSTALL_MESSAGE
-            )
-
-        # Check if Axolotl's cce fork is installed
-        try:
-            from cut_cross_entropy.transformers.patch import AXOLOTL_CCE_FORK
-
-            if not AXOLOTL_CCE_FORK:
-                raise ImportError
-        except ImportError as e:
-            raise ImportError(
-                "Axolotl's fork of cut_cross_entropy is not installed. "
-                + _CCE_INSTALL_MESSAGE
-            ) from e
 
     def pre_model_load(self, cfg):
         """Apply cut cross entropy before model loading if enabled."""
         if cfg.cut_cross_entropy:
             self._check_requirements()
 
-            from cut_cross_entropy.transformers.patch import cce_patch
+            from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
+                cce_patch,
+            )
 
             LOG.info(
                 f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"

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

@@ -0,0 +1,191 @@
+"""Cohere and Cohere2 CCE patch."""
+
+# This patch is based off transformers 4.50.0.
+# It patches the forward function for CohereForCausalLM and Cohere2ForCausalLM.
+# It scales the hidden states by the logit scale in advance instead of the logits as the
+# operation is done internally and should be mathematically equivalent.
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.cohere.modeling_cohere import (
+    KwargsForCausalLM,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >> from transformers import AutoTokenizer, CohereForCausalLM
+
+    >> model = CohereForCausalLM.from_pretrained("CohereForAI/c4ai-command-r-v01")
+    >> tokenizer = AutoTokenizer.from_pretrained("CohereForAI/c4ai-command-r-v01")
+
+    >> prompt = "Hey, are you conscious? Can you talk to me?"
+    >> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >> # Generate
+    >> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        # scale hidden_states by logit_scale in-place of logits
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :] * self.logit_scale,
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = logits * self.logit_scale  # main diff from Llama
+
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def patch_cohere(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.cohere import modeling_cohere
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_cohere.CohereForCausalLM
+        ), f"Expected a CohereForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_cohere.CohereForCausalLM.forward = cce_forward
+    return None
+
+
+def patch_cohere2(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.cohere2 import modeling_cohere2
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_cohere2.Cohere2ForCausalLM
+        ), f"Expected a Cohere2ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_cohere2.Cohere2ForCausalLM.forward = cce_forward
+    return None

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

@@ -0,0 +1,165 @@
+"""Gemma CCE patch"""
+
+# This patch is based off transformers 4.50.0.
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.gemma.modeling_gemma import (
+    KwargsForCausalLM,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, GemmaForCausalLM
+
+    >>> model = GemmaForCausalLM.from_pretrained("google/gemma-7b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "What is your favorite condiment?"
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def patch_gemma(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.gemma import modeling_gemma
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_gemma.GemmaForCausalLM
+        ), f"Expected a GemmaForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_gemma.GemmaForCausalLM.forward = cce_forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/gemma3.py

@@ -0,0 +1,447 @@
+"""Gemma2 and Gemma3 (text and multimodal) CCE patch."""
+
+# Implementation originally adapted from https://github.com/apple/ml-cross-entropy/pull/29
+# and updated for transformers 4.50.0.
+# This is a modified version of the patch that allows for deferred logits calculation for gemma3 and works
+# with both gemma3 (text and multimodal) models.
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+from torch import nn
+from transformers.cache_utils import Cache, HybridCache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.gemma3.modeling_gemma3 import (
+    Gemma3CausalLMOutputWithPast,
+    logger,
+)
+from transformers.utils import (
+    is_torchdynamo_compiling,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+from axolotl.integrations.cut_cross_entropy.monkeypatch.utils import apply_lce
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[HybridCache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    defer_logits_calculation: bool = False,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        defer_logits_calculation (`bool`, *optional*):
+            If `True`, defer logits calculation to the ConditionalGeneration forward. This is used to avoid the
+            memory overhead of calculating logits using regular lm_head forward pass and to use CCE.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Gemma3ForCausalLM
+
+    >>> model = Gemma3ForCausalLM.from_pretrained("google/gemma-2-9b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "What is your favorite condiment?"
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **loss_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            softcap=getattr(self.config, "final_logit_softcapping", None),
+            **loss_kwargs,
+        )
+    elif _PATCH_OPTS is not None and defer_logits_calculation:
+        # defer logits calculation to the ConditionalGeneration forward
+        logits = hidden_states[:, slice_indices, :]
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if self.config.final_logit_softcapping is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward_multimodal(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    pixel_values: torch.FloatTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[list[torch.FloatTensor], Cache]] = None,
+    token_type_ids: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **lm_kwargs,
+) -> Union[Tuple, Gemma3CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.text_config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.text_config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
+
+    >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/Gemma3-test-224px-hf")
+    >>> processor = AutoProcessor.from_pretrained("google/Gemma3-test-224px-hf")
+
+    >>> prompt = "answer en Where is the cow standing?"
+    >>> url = "https://huggingface.co/gv-hf/Gemma3-test-224px-hf/resolve/main/cow_beach_1.png"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_length=30)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "answer en Where is the cow standing?\nbeach"
+    ```"""
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    is_training = token_type_ids is not None and labels is not None
+
+    # Replace image id woth PAD if the image token if OOV, to avoid index-errors
+    if input_ids is not None and self.config.image_token_index >= self.vocab_size:
+        special_image_mask = input_ids == self.config.image_token_index
+        llm_input_ids = input_ids.clone()
+        llm_input_ids[special_image_mask] = 0
+    else:
+        llm_input_ids = input_ids  # type: ignore
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(llm_input_ids)
+
+    if cache_position is None:
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0  # type: ignore
+        )
+        cache_position = torch.arange(  # type: ignore
+            past_seen_tokens,
+            past_seen_tokens + inputs_embeds.shape[1],
+            device=inputs_embeds.device,
+        )
+
+    # Merge text and images
+    if pixel_values is not None:
+        image_features = self.get_image_features(pixel_values)
+
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(
+                    self.config.image_token_index,
+                    dtype=torch.long,
+                    device=inputs_embeds.device,
+                )
+            )
+        else:
+            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(
+                -1
+            )
+            special_image_mask = special_image_mask.expand_as(inputs_embeds).to(
+                inputs_embeds.device
+            )
+
+        if (
+            not is_torchdynamo_compiling()
+            and inputs_embeds[special_image_mask].numel() != image_features.numel()
+        ):
+            image_tokens_in_text = (special_image_mask).sum(dim=1).sum(dim=0)[0]
+            raise ValueError(
+                f"Number of images does not match number of special image tokens in the input text. "
+                f"Got {image_tokens_in_text} image tokens in the text but {image_features.shape[0] * image_features.shape[1]} "
+                "tokens from image embeddings."
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)  # type: ignore
+
+    # mask out pad-token-ids in labels for BC
+    if labels is not None and self.pad_token_id in labels:
+        logger.warning_once(
+            "`labels` contains `pad_token_id` which will be masked with `config.ignore_index`. "
+            "You have to mask out `pad_token_id` when preparing `labels`, this behavior will be removed in v.4.46.",
+        )
+        labels = torch.where(  # type: ignore
+            input_ids == self.pad_token_id, self.config.ignore_index, labels
+        )
+
+    causal_mask = self._update_causal_mask(  # pylint: disable=protected-access
+        attention_mask,
+        token_type_ids,
+        past_key_values,
+        cache_position,
+        inputs_embeds,
+        is_training,
+    )
+    outputs = self.language_model(
+        attention_mask=causal_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        defer_logits_calculation=True,  # enable deferred logits calculation
+        **lm_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.language_model.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            softcap=getattr(self.config, "final_logit_softcapping", None),
+            **lm_kwargs,
+        )
+    else:
+        logits = hidden_states
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            shift_logits = logits[..., :-1, :]
+            shift_labels = labels[..., 1:]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(
+                    logits.device
+                )
+                shift_logits = shift_logits[
+                    shift_attention_mask.to(logits.device) != 0
+                ].contiguous()
+                shift_labels = shift_labels[
+                    shift_attention_mask.to(shift_labels.device) != 0
+                ].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Gemma3CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def patch_gemma2(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.gemma2 import modeling_gemma2
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_gemma2.Gemma2ForCausalLM
+        ), f"Expected a Gemma2ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_gemma2.Gemma2ForCausalLM.forward = cce_forward
+    return None
+
+
+def patch_gemma3_text(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.gemma3 import modeling_gemma3
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_gemma3.Gemma3ForCausalLM
+        ), f"Expected a Gemma3ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_gemma3.Gemma3ForCausalLM.forward = cce_forward
+    return None
+
+
+def patch_gemma3(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.gemma3 import modeling_gemma3
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_gemma3.Gemma3ForConditionalGeneration
+        ), f"Expected a Gemma3ForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        # patch the causal model to enable deferred logits calculation
+        maybe_model.language_model.forward = MethodType(
+            cce_forward, maybe_model.language_model
+        )
+        return maybe_model
+
+    modeling_gemma3.Gemma3ForConditionalGeneration.forward = cce_forward_multimodal
+    # patch the causal model to enable deferred logits calculation
+    modeling_gemma3.Gemma3ForCausalLM.forward = cce_forward
+    return None

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/llama.py

@@ -0,0 +1,164 @@
+"""Llama CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.models.llama.modeling_llama import (
+    KwargsForCausalLM,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> CausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+    >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: BaseModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def patch_llama(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    """Patch Llama for CCE."""
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.llama import modeling_llama
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_llama.LlamaForCausalLM
+        ), f"Expected a LlamaForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_llama.LlamaForCausalLM.forward = cce_forward
+    return None

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

@@ -0,0 +1,401 @@
+"""Llama4 CCE patch. Adapted from transformers 4.51.0."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch import nn
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.llama4.modeling_llama4 import (
+    Llama4CausalLMOutputWithPast,
+)
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    defer_logits_calculation: bool = False,
+    **kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        defer_logits_calculation (`bool`, *optional*, defaults to `False`):
+            If `True`, defer logits calculation to the ConditionalGeneration forward. This is used to avoid the
+            memory overhead of calculating logits using regular lm_head forward pass and to use CCE.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Llama4ForCausalLM
+
+    >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    elif _PATCH_OPTS is not None and defer_logits_calculation:
+        # defer logits calculation to the ConditionalGeneration forward
+        logits = hidden_states[:, slice_indices, :]
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def cce_forward_multimodal(
+    self,
+    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,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[Union[int, list[int]]] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: torch.Tensor | None = None,
+    **lm_kwargs,
+) -> Union[Tuple, Llama4CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+    >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
+    >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+
+    >>> prompt = "USER: <image>\nWhat's the content of the image? ASSISTANT:"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "USER:  \nWhat's the content of the image? ASSISTANT: The image features a busy city street with a stop sign prominently displayed"
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+    vision_feature_layer = (
+        vision_feature_layer
+        if vision_feature_layer is not None
+        else self.config.vision_config.vision_feature_layer
+    )
+    vision_feature_select_strategy = (
+        vision_feature_select_strategy
+        if vision_feature_select_strategy is not None
+        else self.config.vision_config.vision_feature_select_strategy
+    )
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    if pixel_values is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+        )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(input_ids)  # type: ignore
+
+    if pixel_values is not None:
+        image_features = self.get_image_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            vision_feature_select_strategy=vision_feature_select_strategy,
+            image_sizes=image_sizes,
+        )
+        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)  # type: ignore
+        inputs_embeds = inputs_embeds.view(-1, inputs_embeds.size(-1))  # type: ignore
+
+        final_mask_1d = final_mask[..., 0].reshape(-1)
+        num_tokens_to_fill = final_mask_1d.sum()
+
+        if num_tokens_to_fill != projected_vision_flat.size(0):
+            raise ValueError(
+                f"Mismatch: final_mask wants {num_tokens_to_fill} embeddings, "
+                f"but multi_modal_projector returned {projected_vision_flat.size(0)}"
+            )
+
+        expanded_mask = final_mask_1d.unsqueeze(-1).expand(-1, inputs_embeds.size(-1))
+        inputs_embeds = inputs_embeds.masked_scatter(
+            expanded_mask, projected_vision_flat
+        )  # type: ignore
+        inputs_embeds = inputs_embeds.view(original_inputs_embeds_shape)  # type: ignore
+
+    outputs = self.language_model(
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        defer_logits_calculation=True,  # enable deferred logits calculation
+        **lm_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        # TODO: check if need to handle attention_mask
+        loss = apply_lce(
+            hidden_states,
+            self.language_model.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **lm_kwargs,
+        )
+    else:
+        logits = hidden_states
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(
+                    logits.device
+                )
+                shift_logits = logits[..., :-1, :][
+                    shift_attention_mask.to(logits.device) != 0
+                ].contiguous()
+                shift_labels = labels[..., 1:][
+                    shift_attention_mask.to(labels.device) != 0
+                ].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1).to(shift_logits.device),
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Llama4CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,  # type: ignore  # TODO: check if need to create dummy logits
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def patch_llama4_text(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.llama4 import modeling_llama4
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_llama4.Llama4ForCausalLM
+        ), f"Expected a Llama4ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+
+        return maybe_model
+
+    setattr(
+        modeling_llama4.Llama4ForCausalLM,
+        "forward",
+        cce_forward,
+    )
+    return None
+
+
+def patch_llama4(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.llama4 import modeling_llama4
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_llama4.Llama4ForConditionalGeneration
+        ), f"Expected a Llama4ForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        # patch the language model
+        maybe_model.language_model.forward = MethodType(
+            cce_forward, maybe_model.language_model
+        )
+        return maybe_model
+
+    setattr(
+        modeling_llama4.Llama4ForConditionalGeneration,
+        "forward",
+        cce_forward_multimodal,
+    )
+
+    # patch the causal language model
+    setattr(modeling_llama4.Llama4ForCausalLM, "forward", cce_forward)
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/mistral3.py

@@ -0,0 +1,384 @@
+"""Mistral and Mistral3 CCE patch."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch import nn
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.mistral3.modeling_mistral3 import (
+    Mistral3CausalLMOutputWithPast,
+)
+from transformers.models.mistral.modeling_mistral import (
+    KwargsForCausalLM,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    is_torchdynamo_compiling,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] | None = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    defer_logits_calculation: bool = False,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        defer_logits_calculation (`bool`, *optional*):
+            If `True`, defer logits calculation to the ConditionalGeneration forward. This is used to avoid the
+            memory overhead of calculating logits using regular lm_head forward pass and to use CCE.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, MistralForCausalLM
+
+    >>> model = MistralForCausalLM.from_pretrained("meta-mistral/Mistral-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-mistral/Mistral-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    elif _PATCH_OPTS is not None and defer_logits_calculation:
+        # defer logits calculation to the ConditionalGeneration forward
+        logits = hidden_states[:, slice_indices, :]
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def cce_forward_multimodal(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    pixel_values: torch.FloatTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[Union[int, list[int]]] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    image_sizes: torch.Tensor | None = None,
+    **lm_kwargs,
+) -> Union[Tuple, Mistral3CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Mistral3ForConditionalGeneration
+
+    >>> model = Mistral3ForConditionalGeneration.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")
+    >>> processor = AutoProcessor.from_pretrained("mistralai/Mistral-Small-3.1-24B-Instruct-2503")
+
+    >>> prompt = "<s>[INST][IMG]What is the image?[/INST]"
+    >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_new_tokens=15)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "What is the image?The image depicts two cats lying on a pink blanket."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+    vision_feature_layer = (
+        vision_feature_layer
+        if vision_feature_layer is not None
+        else self.config.vision_feature_layer
+    )
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    if pixel_values is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+        )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+
+    if pixel_values is not None:
+        image_features = self.get_image_features(
+            pixel_values=pixel_values,
+            vision_feature_layer=vision_feature_layer,
+            image_sizes=image_sizes,
+        )
+
+        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(
+            inputs_embeds.device
+        )
+        if (
+            not is_torchdynamo_compiling()
+            and inputs_embeds[special_image_mask].numel() != image_features.numel()
+        ):
+            n_image_tokens = (input_ids == self.config.image_token_index).sum()
+            n_image_features = image_features.shape[0] * image_features.shape[1]
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)  # type: ignore
+
+    outputs = self.language_model(
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        defer_logits_calculation=True,  # enable deferred logits calculation
+        **lm_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.language_model.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **lm_kwargs,
+        )
+    else:
+        logits = hidden_states
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -(logits.shape[1] - 1) :].to(
+                    logits.device
+                )
+                shift_logits = logits[..., :-1, :][
+                    shift_attention_mask.to(logits.device) != 0
+                ].contiguous()
+                shift_labels = labels[..., 1:][
+                    shift_attention_mask.to(labels.device) != 0
+                ].contiguous()
+            else:
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)),
+                shift_labels.view(-1).to(shift_logits.device),
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Mistral3CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def patch_mistral(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.mistral import modeling_mistral
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_mistral.MistralForCausalLM
+        ), f"Expected a MistralForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_mistral.MistralForCausalLM.forward = cce_forward
+    return None
+
+
+def patch_mistral3(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.mistral import modeling_mistral
+    from transformers.models.mistral3 import modeling_mistral3
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_mistral3.Mistral3ForConditionalGeneration
+        ), f"Expected a Mistral3ForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        # patch the causal model to enable deferred logits calculation
+        maybe_model.language_model.forward = MethodType(
+            cce_forward, maybe_model.language_model
+        )
+        return maybe_model
+
+    modeling_mistral3.Mistral3ForConditionalGeneration.forward = cce_forward_multimodal
+    # patch the causal model to enable deferred logits calculation
+    modeling_mistral.MistralForCausalLM.forward = cce_forward
+    return None

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

@@ -0,0 +1,366 @@
+"""Mllama CCE patch."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.mllama.modeling_mllama import (
+    _prepare_cross_attention_mask,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: torch.LongTensor | None = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    cross_attention_states: Optional[torch.LongTensor] = None,
+    cross_attention_mask: Optional[torch.LongTensor] = None,
+    full_text_row_masked_out_mask: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    defer_logits_calculation: bool = False,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        defer_logits_calculation (`bool`, *optional*):
+            If `True`, defer logits calculation to the ConditionalGeneration forward. This is used to avoid the
+            memory overhead of calculating logits using regular lm_head forward pass and to use CCE.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, MllamaForCausalLM
+
+    >>> model = MllamaForCausalLM.from_pretrained("Llama-3.2-11B-Vision")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Llama-3.2-11B-Vision")
+
+    >>> prompt = "If I had to write a haiku, it would be:"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=40, do_sample=True, temperature=0.6)
+    >>> result = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    >>> print(result)
+    If I had to write a haiku, it would be: "Snowflakes gently fall" - simple, yet peaceful.
+    I love the idea of snowflakes gently falling, each one
+    ```
+    """
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        cross_attention_states=cross_attention_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        cross_attention_mask=cross_attention_mask,
+        full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **loss_kwargs,
+        )
+    elif _PATCH_OPTS is not None and defer_logits_calculation:
+        # defer logits calculation to the ConditionalGeneration forward
+        logits = hidden_states[:, slice_indices, :]
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :]).float()
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward_multimodal(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    aspect_ratio_mask: Optional[torch.Tensor] = None,
+    aspect_ratio_ids: Optional[torch.Tensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    cross_attention_mask: Optional[torch.Tensor] = None,
+    cross_attention_states: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, MllamaForConditionalGeneration
+
+    >>> checkpoint = "meta-llama/Llama-3.2-11B-Vision"
+    >>> model = MllamaForConditionalGeneration.from_pretrained(checkpoint)
+    >>> processor = AutoProcessor.from_pretrained(checkpoint)
+
+    >>> prompt = "<|image|>If I had to write a haiku for this one"
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(text=prompt, images=image, return_tensors="pt")
+
+    >>> # Generate
+    >>> output = model.generate(**inputs, max_new_tokens=15)
+
+    >>> prompt_len = inputs.input_ids.shape[-1]
+    >>> generated_ids = output[:, prompt_len:]
+    >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    >>> print(generated_text)
+    [', it would be:.\\nA stop sign in Chinatown.\\n']
+    ```
+    """
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    if pixel_values is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+        )
+
+    if pixel_values is not None and cross_attention_states is not None:
+        raise ValueError(
+            "`pixel_values` and `cross_attention_states` cannot be provided simultaneously"
+        )
+
+    if pixel_values is not None:
+        if aspect_ratio_ids is None:
+            raise ValueError(
+                "`aspect_ratio_ids` must be provided if `pixel_values` is provided"
+            )
+        # get vision tokens from vision model
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            aspect_ratio_ids=aspect_ratio_ids,
+            aspect_ratio_mask=aspect_ratio_mask,
+            output_hidden_states=output_hidden_states,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+        cross_attention_states = vision_outputs[0]
+        cross_attention_states = self.multi_modal_projector(
+            cross_attention_states
+        ).reshape(
+            -1, cross_attention_states.shape[-2], self.hidden_size  # type: ignore
+        )
+
+    if cross_attention_mask is not None:
+        cross_attention_mask, full_text_row_masked_out_mask = (
+            _prepare_cross_attention_mask(
+                cross_attention_mask,
+                num_vision_tokens=self.vision_model.num_patches,
+                dtype=self.dtype,
+            )
+        )
+    else:
+        full_text_row_masked_out_mask = None
+
+    if cross_attention_mask is not None and cache_position is not None:
+        cross_attention_mask = cross_attention_mask[:, :, cache_position]
+        full_text_row_masked_out_mask = full_text_row_masked_out_mask[
+            :, :, cache_position
+        ]
+
+    outputs = self.language_model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        cross_attention_states=cross_attention_states,
+        cross_attention_mask=cross_attention_mask,
+        full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+        past_key_values=past_key_values,
+        use_cache=use_cache,
+        inputs_embeds=inputs_embeds,
+        output_hidden_states=output_hidden_states,
+        output_attentions=output_attentions,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        defer_logits_calculation=True,  # enable deferred logits calculation
+        **loss_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    loss = None
+    logits = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.language_model.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **loss_kwargs,
+        )
+    else:
+        # Temporary fix to calculate the loss in main class, as the model's vocab size may be resized
+        logits = hidden_states
+
+        if labels is not None:
+            loss = self.loss_function(
+                logits, labels, self.config.get_text_config().vocab_size, **loss_kwargs
+            )
+
+    if not return_dict:
+        return (loss,) + outputs if loss is not None else outputs
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=outputs.logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def patch_mllama(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.mllama import modeling_mllama
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_mllama.MllamaForConditionalGeneration
+        ), f"Expected a MllamaForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        # patch the language model
+        maybe_model.language_model.forward = MethodType(
+            cce_forward, maybe_model.language_model
+        )
+        return maybe_model
+
+    modeling_mllama.MllamaForConditionalGeneration.forward = cce_forward_multimodal
+
+    # patch the causal language model
+    modeling_mllama.MllamaForCausalLM.forward = cce_forward
+    return None

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

@@ -0,0 +1,126 @@
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+
+"""Cut Cross Entropy patcher"""
+
+import transformers
+from cut_cross_entropy.cce_utils import LinearCrossEntropyImpl
+from cut_cross_entropy.linear_cross_entropy import LCE_IMPL_DEFAULT
+from cut_cross_entropy.transformers.phi3 import patch_phi3
+from cut_cross_entropy.transformers.utils import PatchOptions, TransformersModelT
+
+from axolotl.integrations.cut_cross_entropy.monkeypatch.cohere import (
+    patch_cohere,
+    patch_cohere2,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.gemma import patch_gemma
+from axolotl.integrations.cut_cross_entropy.monkeypatch.gemma3 import (
+    patch_gemma2,
+    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.llama import (
+    patch_llama,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.llama4 import (
+    patch_llama4,
+    patch_llama4_text,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.mistral3 import (
+    patch_mistral,
+    patch_mistral3,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.mllama import patch_mllama
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2 import (
+    patch_qwen2,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_5_vl import (
+    patch_qwen2_5_vl,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_moe import (
+    patch_qwen2_moe,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_vl import (
+    patch_qwen2_vl,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen3 import patch_qwen3
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen3_moe import (
+    patch_qwen3_moe,
+)
+
+CUT_CROSS_ENTROPY_MODEL_MAPPING = {
+    "llama": patch_llama,
+    "llama4": patch_llama4,
+    "llama4_text": patch_llama4_text,
+    "mllama": patch_mllama,
+    "phi3": patch_phi3,
+    "gemma": patch_gemma,
+    "gemma2": patch_gemma2,
+    "gemma3": patch_gemma3,
+    "gemma3_text": patch_gemma3_text,
+    "mistral": patch_mistral,
+    "mistral3": patch_mistral3,
+    "qwen2": patch_qwen2,
+    "qwen2_moe": patch_qwen2_moe,
+    "qwen2_vl": patch_qwen2_vl,
+    "qwen2_5_vl": patch_qwen2_5_vl,
+    "qwen3": patch_qwen3,
+    "qwen3_moe": patch_qwen3_moe,
+    "cohere": patch_cohere,
+    "cohere2": patch_cohere2,
+    "glm": patch_glm,
+    "glm4": patch_glm4,
+}
+
+
+def cce_patch(
+    model_type_or_model: str | TransformersModelT | transformers.PretrainedConfig,
+    impl: str | LinearCrossEntropyImpl = LCE_IMPL_DEFAULT,
+    reduction: str = "mean",
+    filter_eps: float | str | None = "auto",
+    accum_e_fp32: bool = False,
+    accum_c_fp32: bool = False,
+    filter_e_grad: bool = True,
+    filter_c_grad: bool = True,
+    train_only: bool = False,
+) -> TransformersModelT | None:
+    if isinstance(impl, LinearCrossEntropyImpl):
+        impl = impl.name.lower()
+
+    if impl not in (v.name.lower() for v in LinearCrossEntropyImpl):
+        raise ValueError(f"Unknown {impl=}")
+
+    if isinstance(model_type_or_model, transformers.PreTrainedModel):
+        if hasattr(model_type_or_model, "config"):
+            model_type = getattr(
+                getattr(model_type_or_model, "config", None), "model_type", None
+            )
+        else:
+            raise ValueError(
+                "model_type_or_model is a PreTrainedModel but does not have a config attribute"
+            )
+    elif isinstance(model_type_or_model, transformers.PretrainedConfig):
+        model_type = model_type_or_model.model_type
+    else:
+        model_type = model_type_or_model
+
+    patch_options = PatchOptions(
+        impl=impl,
+        reduction=reduction,
+        filter_eps=filter_eps,
+        accum_e_fp32=accum_e_fp32,
+        accum_c_fp32=accum_c_fp32,
+        filter_e_grad=filter_e_grad,
+        filter_c_grad=filter_c_grad,
+        train_only=train_only,
+    )
+
+    if model_type in CUT_CROSS_ENTROPY_MODEL_MAPPING:
+        return CUT_CROSS_ENTROPY_MODEL_MAPPING[model_type](
+            model_type_or_model, patch_options
+        )
+
+    raise RuntimeError(f"Unknown model type {model_type}")

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2.py

@@ -0,0 +1,37 @@
+"""Qwen2 CCE patch. The model inherits Llama's modeling code and uses the same forward method."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+
+
+def patch_qwen2(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    from transformers.models.qwen2 import modeling_qwen2
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import axolotl.integrations.cut_cross_entropy.monkeypatch.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from axolotl.integrations.cut_cross_entropy.monkeypatch.llama import (
+        cce_forward,
+    )
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2.Qwen2ForCausalLM
+        ), f"Expected a Qwen2ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_qwen2.Qwen2ForCausalLM.forward = cce_forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_5_vl.py

@@ -0,0 +1,246 @@
+"""Qwen2.5 VL CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2_5_VLCausalLMOutputWithPast,
+)
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+def cce_forward_multimodal(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            image_mask = mask_expanded.to(inputs_embeds.device)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)  # type: ignore
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            video_mask = mask_expanded.to(inputs_embeds.device)
+
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)  # type: ignore
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (
+            (cache_position is not None and cache_position[0] == 0)
+            or self.rope_deltas is None
+            or (past_key_values is None or past_key_values.get_seq_length() == 0)  # type: ignore
+        ):
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                attention_mask,
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (
+                (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
+                if cache_position is not None
+                else 0
+            )
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)  # type: ignore
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)  # type: ignore
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)  # type: ignore
+            position_ids = position_ids.add(delta)  # type: ignore
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)  # type: ignore
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    logits = None
+    loss = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2_5_VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=self.rope_deltas,
+    )
+
+
+def patch_qwen2_5_vl(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_5_vl import modeling_qwen2_5_vl
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration
+        ), f"Expected a Qwen2_5_VLForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = (
+        cce_forward_multimodal
+    )
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_moe.py

@@ -0,0 +1,178 @@
+"""Qwen2 MoE CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.models.qwen2_moe.modeling_qwen2_moe import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    load_balancing_loss_func,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen2MoeForCausalLM
+
+    >>> model = Qwen2MoeForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+    >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    loss = None
+    logits = None
+
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **loss_kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(  # type: ignore
+                loss.device  # type: ignore
+            )  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,  # type: ignore
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )
+
+
+def patch_qwen2_moe(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_moe import modeling_qwen2_moe
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_moe.Qwen2MoeForCausalLM
+        ), f"Expected a Qwen3MoeForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(forward, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_moe.Qwen2MoeForCausalLM.forward = forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_vl.py

@@ -0,0 +1,239 @@
+"""Qwen2 VL CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+    Qwen2VLCausalLMOutputWithPast,
+)
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+def cce_forward_multimodal(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+    >>> model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.get_dtype())
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_mask = (
+                (input_ids == self.config.image_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
+            )
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)  # type: ignore
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.get_dtype())
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+            video_mask = (
+                (input_ids == self.config.video_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
+            )
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)  # type: ignore
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (
+            (cache_position is not None and cache_position[0] == 0)
+            or self.rope_deltas is None
+            or (past_key_values is None or past_key_values.get_seq_length() == 0)  # type: ignore
+        ):
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids, image_grid_thw, video_grid_thw, attention_mask
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (
+                cache_position[0] + self.rope_deltas
+                if cache_position is not None
+                else 0
+            )
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)  # type: ignore
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)  # type: ignore
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)  # type: ignore
+                delta = delta.to(position_ids.device)  # type: ignore
+            position_ids = position_ids.add(delta)  # type: ignore
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)  # type: ignore
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    logits = None
+    loss = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=self.rope_deltas,
+    )
+
+
+def patch_qwen2_vl(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_vl import modeling_qwen2_vl
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_vl.Qwen2VLForConditionalGeneration
+        ), f"Expected a Qwen2VLForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_vl.Qwen2VLForConditionalGeneration.forward = cce_forward_multimodal
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen3.py

@@ -0,0 +1,35 @@
+"""Qwen3 CCE patch. The model inherits Llama's modeling code and uses the same forward method."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+
+
+def patch_qwen3(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    from transformers.models.qwen3 import modeling_qwen3
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import axolotl.integrations.cut_cross_entropy.monkeypatch.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from axolotl.integrations.cut_cross_entropy.monkeypatch.llama import cce_forward
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen3.Qwen3ForCausalLM
+        ), f"Expected a Qwen3ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_qwen3.Qwen3ForCausalLM.forward = cce_forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen3_moe.py

@@ -0,0 +1,183 @@
+"""Qwen3 MoE CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.models.qwen3_moe.modeling_qwen3_moe import (
+    KwargsForCausalLM,
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    load_balancing_loss_func,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
+
+    >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(  # type: ignore
+                loss.device  # type: ignore
+            )  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,  # type: ignore
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )
+
+
+def patch_qwen3_moe(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen3_moe.Qwen3MoeForCausalLM
+        ), f"Expected a Qwen3MoeForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(forward, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen3_moe.Qwen3MoeForCausalLM.forward = forward
+    return None

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

@@ -0,0 +1,40 @@
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+
+"""Monkeypatch for apply_lce to add softcap."""
+
+import torch
+from cut_cross_entropy import linear_cross_entropy
+from cut_cross_entropy.transformers.utils import PatchOptions
+
+
+def apply_lce(
+    e: torch.Tensor,
+    c: torch.Tensor,
+    labels: torch.Tensor,
+    opts: PatchOptions,
+    bias: torch.Tensor | None = None,
+    softcap: float | None = None,
+    **loss_kwargs,
+) -> torch.Tensor:
+    """Monkey patch for apply_lce to support softcap kwarg."""
+    num_items_in_batch = loss_kwargs.get("num_items_in_batch", None)
+    cce_kwargs = opts.to_kwargs()
+    if num_items_in_batch is not None and cce_kwargs["reduction"] == "mean":
+        cce_kwargs["reduction"] = "sum"
+    else:
+        num_items_in_batch = None
+
+    loss = linear_cross_entropy(
+        e,
+        c,
+        labels.to(e.device),
+        bias=bias,
+        shift=True,
+        softcap=softcap,
+        **cce_kwargs,
+    )
+
+    if num_items_in_batch is not None:
+        loss = loss / num_items_in_batch
+
+    return loss

src/axolotl/integrations/liger/__init__.py

@@ -27,7 +27,7 @@ from axolotl.utils.logging import get_logger
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 from .utils import patch_with_compile_disable
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 
 class LigerPlugin(BasePlugin):

src/axolotl/integrations/liger/args.py

@@ -15,7 +15,6 @@
 """
 Module for handling LIGER input arguments.
 """
-
 from typing import Optional
 
 from pydantic import BaseModel, model_validator

src/axolotl/loaders/model.py

@@ -504,9 +504,6 @@ class ModelLoader:
                 # for some reason, this causes the loss to be off by an order of magnitude
                 # but deepspeed needs this still in bfloat16
                 bnb_config["bnb_4bit_quant_storage"] = torch.float32
-            if self.cfg.model_config_type == "falcon_h1":
-                # output projection cannot be quantized for Falcon-H1 models
-                bnb_config["llm_int8_skip_modules"] = ["out_proj"]
 
             if self.cfg.bnb_config_kwargs:
                 bnb_config.update(self.cfg.bnb_config_kwargs)
@@ -521,9 +518,6 @@ class ModelLoader:
             # Exclude mamba blocks from int8 quantization for jamba
             if self.cfg.model_config_type == "jamba":
                 bnb_config["llm_int8_skip_modules"] = ["mamba"]
-            if self.cfg.model_config_type == "falcon_h1":
-                # output projection cannot be quantized for Falcon-H1 models
-                bnb_config["llm_int8_skip_modules"] = ["out_proj"]
             self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                 **bnb_config,
             )

src/axolotl/loaders/patch_manager.py

@@ -50,7 +50,6 @@ class PatchManager:
     def apply_pre_model_load_patches(self):
         """Apply pre-model load patches based on config."""
         self._apply_flash_attention_patches()
-        self._apply_chunked_cross_entropy_patch()
         self._apply_fsdp_patches()
         self._apply_adapter_patches()
         self._apply_flex_attention_patches()
@@ -64,7 +63,6 @@ class PatchManager:
         self._patch_llama_derived_model()
         self._apply_mistral_cross_entropy_patch()
         self._apply_self_attention_lora_patch()
-        self._apply_gemma3_conditional_generation_forward_patch()
 
     def apply_post_model_load_patches(self, model: PreTrainedModel):
         """Apply patches that require the model instance."""
@@ -80,15 +78,6 @@ class PatchManager:
             patch_xformers_attn_over_fa2()
             self.cfg.flash_attention = True
 
-    def _apply_chunked_cross_entropy_patch(self):
-        if self.cfg.chunked_cross_entropy:
-            from axolotl.monkeypatch.loss.chunked import patch_chunked_ce_loss_fn
-
-            if self.cfg.chunked_cross_entropy_num_chunks:
-                patch_chunked_ce_loss_fn(self.cfg.chunked_cross_entropy_num_chunks)
-            else:
-                patch_chunked_ce_loss_fn()
-
     def _apply_fsdp_patches(self):
         """Apply patches for FSDP configurations."""
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
@@ -222,15 +211,6 @@ class PatchManager:
                 has_remote_code=has_remote_code,
             )
 
-    def _apply_gemma3_conditional_generation_forward_patch(self):
-        """Apply gemma3 conditional generation forward patch."""
-        if self.model_config.model_type in ["gemma3", "gemma3_text"]:
-            from axolotl.monkeypatch.models.gemma3.modeling import (
-                patch_gemma3_conditional_generation_forward,
-            )
-
-            patch_gemma3_conditional_generation_forward()
-
     def _patch_attention(self):
         """Apply attention-specific patches based on model type."""
         if not (self.cfg.flash_attention and hasattr(self.model_config, "model_type")):

src/axolotl/loaders/tokenizer.py

@@ -273,7 +273,7 @@ def load_tokenizer(cfg: DictDefault) -> PreTrainedTokenizer:
             {"additional_special_tokens": additional_special_tokens}
         )
 
-    if is_main_process():
+    if is_main_process(use_environ=True):
         LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
         LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
         LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")

src/axolotl/monkeypatch/data/batch_dataset_fetcher.py

@@ -9,9 +9,6 @@ from torch.utils.data._utils.worker import _worker_loop
 
 class _MapDatasetFetcher(_BaseDatasetFetcher):
     def fetch(self, possibly_batched_index):
-        if not possibly_batched_index:
-            return self.collate_fn([])
-
         if isinstance(possibly_batched_index[0], list):
             data = [None for i in possibly_batched_index]
             for i, possibly_batched_index_ in enumerate(possibly_batched_index):

src/axolotl/monkeypatch/loss/chunked.py

@@ -1,134 +0,0 @@
-"""
-chunked ce loss
-"""
-
-from typing import List, Optional
-
-import torch
-import torch.nn.functional as F
-
-
-# copied and modified from torchtune.modules.loss.CEWithChunkedOutputLoss
-class CEWithChunkedOutputLoss(torch.nn.Module):
-    """
-    Cross-entropy with chunked outputs that saves memory by only upcasting one chunk at a time.
-
-    For more details, please refer to: https://github.com/pytorch/torchtune/pull/1390
-    """
-
-    def __init__(self, num_output_chunks: int = 8, ignore_index: int = -100):
-        super().__init__()
-        self.num_output_chunks = num_output_chunks
-        self.ignore_index = ignore_index
-
-    def compute_cross_entropy(
-        self,
-        logits: torch.Tensor,
-        labels: torch.Tensor,
-        normalize: bool = True,  # pylint: disable=unused-argument
-    ) -> torch.Tensor:
-        """
-        Upcast logits to fp32 and compute cross entropy loss.
-        """
-        return F.cross_entropy(
-            logits.float(), labels, ignore_index=self.ignore_index, reduction="sum"
-        )
-
-    def forward(
-        self, logits: List[torch.Tensor], labels: torch.Tensor, reduction="sum"
-    ) -> torch.Tensor:
-        """
-        Args:
-            logits (List[torch.Tensor]): List of chunked logits of length
-                ``self.num_output_chunks``, where each chunk has shape
-                ``(batch_size, num_tokens / num_output_chunks, vocab_size)``.
-            labels (torch.Tensor): Ground truth labels of shape ``(batch_size, num_tokens)``.
-            reduction (str): The reduction to apply to the output.
-
-        Returns:
-            torch.Tensor: Cross entropy loss of shape (1,).
-        """
-
-        total_elements = (labels != self.ignore_index).sum()
-
-        # chunk and reshape labels (bsz, num_tokens, vocab) -> [(bsz*num_tokens/num_chunks, vocab)]
-        labels = [
-            target_chunk.reshape(-1)
-            for target_chunk in labels.chunk(self.num_output_chunks, dim=1)
-        ]
-        # reshape logits [(bsz, num_tokens/num_chunks, vocab)] -> [(bsz*num_tokens/num_chunks, vocab)]
-        logits = [
-            logit_chunk.reshape(-1, logit_chunk.size(-1)) for logit_chunk in logits
-        ]
-
-        # compute one chunk at a time
-        total_loss = 0.0
-        for logits_chunk, labels_chunk in zip(logits, labels):
-            total_loss += self.compute_cross_entropy(logits_chunk, labels_chunk)
-
-        if reduction == "sum":
-            return total_loss
-        return total_loss / total_elements
-
-
-def _build_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
-    loss_fn_ce = CEWithChunkedOutputLoss(num_output_chunks, ignore_index)
-    loss_fn_ce.compute_cross_entropy = torch.compile(
-        loss_fn_ce.compute_cross_entropy, backend="inductor"
-    )
-    return loss_fn_ce
-
-
-def get_causal_lm_loss(num_output_chunks: int = 8, ignore_index: int = -100):
-    loss_fn_ce = _build_chunked_ce_loss_fn(num_output_chunks, ignore_index)
-
-    def chunked_fix_cross_entropy(
-        source,
-        target,
-        num_items_in_batch: int = None,
-        ignore_index: int = -100,
-        **kwargs,
-    ):  # pylint: disable=unused-argument
-        reduction = "sum" if num_items_in_batch is not None else "mean"
-        logit_chunks = [  # pylint: disable=unnecessary-comprehension
-            chunk for chunk in source.chunk(loss_fn_ce.num_output_chunks, dim=1)
-        ]
-        loss = loss_fn_ce(logit_chunks, target, reduction=reduction)
-        if reduction == "sum":
-            loss = loss / num_items_in_batch
-        return loss
-
-    def for_causal_lm_chunked_loss(
-        logits,
-        labels,
-        vocab_size: int = None,  # pylint: disable=unused-argument
-        num_items_in_batch: Optional[int] = None,
-        ignore_index: int = -100,
-        shift_labels: Optional[torch.Tensor] = None,
-        **kwargs,
-    ) -> torch.Tensor:
-        # skip the upcast to float since we handle that in the chunking loss
-        if shift_labels is None:
-            # Shift so that tokens < n predict n
-            labels = F.pad(labels, (0, 1), value=ignore_index)
-            shift_labels = labels[..., 1:].contiguous()
-
-        # Skip Flattening the tokens
-        # Enable model parallelism
-        shift_labels = shift_labels.to(logits.device)
-        loss = chunked_fix_cross_entropy(
-            logits, shift_labels, num_items_in_batch, ignore_index, **kwargs
-        )
-        return loss
-
-    return for_causal_lm_chunked_loss
-
-
-def patch_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
-    import transformers.loss.loss_utils
-
-    for_causal_lm_chunked_loss = get_causal_lm_loss(num_output_chunks, ignore_index)
-    transformers.loss.loss_utils.ForCausalLMLoss = for_causal_lm_chunked_loss
-    transformers.loss.loss_utils.LOSS_MAPPING["ForCausalLM"] = (
-        for_causal_lm_chunked_loss
-    )

src/axolotl/monkeypatch/models/gemma3/modeling.py

@@ -1,16 +0,0 @@
-"""Monkeypatch for gemma3 conditional generation forward to fix high loss"""
-
-
-def patch_gemma3_conditional_generation_forward():
-    # Remove when https://github.com/huggingface/transformers/pull/37208 merged
-
-    from transformers.models.gemma3.modeling_gemma3 import (
-        Gemma3ForConditionalGeneration,
-    )
-
-    setattr(Gemma3ForConditionalGeneration, "accepts_loss_kwargs", False)
-
-    def unpatch():
-        delattr(Gemma3ForConditionalGeneration, "accepts_loss_kwargs")
-
-    return unpatch

src/axolotl/monkeypatch/ring_attn/patch.py

@@ -13,9 +13,9 @@ import inspect
 import accelerate
 import torch
 import torch.distributed as dist
+from accelerate.logging import get_logger
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.enums import RingAttnFunc
 
 LOG = get_logger(__name__)

src/axolotl/monkeypatch/unsloth_.py

@@ -4,12 +4,12 @@ import inspect
 import types
 
 import torch
+from accelerate.logging import get_logger
 from peft import PeftModelForCausalLM
 from torch import nn
 from transformers.models.llama.modeling_llama import LlamaFlashAttention2
 
 from axolotl.monkeypatch.utils import detab_code
-from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 

src/axolotl/processing_strategies.py

@@ -142,7 +142,7 @@ class ProcessingStrategy:
                 # TODO: check if it's normal to be single image only for common datasets
                 # From observation, it's usually a list of single image but some datasets may have several columns for images
                 # Temporary solution: take the first image and suggest people convert their datasets to use multi-content Messages
-                if len(processed_example[image_key]) > 1:
+                if len(processed_example[image_key]) > 0:
                     LOG.warning(
                         f"Found {len(processed_example[image_key])} images in a sample. Using the first one."
                         "If you are using a dataset with multiple images per sample, please convert it to use multi-content Messages."

src/axolotl/train.py

@@ -23,6 +23,7 @@ from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from transformers.trainer import Trainer
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
@@ -218,7 +219,6 @@ def execute_training(
                     gradient_accumulation_steps=cfg.gradient_accumulation_steps,
                     ring_attn_func=cfg.ring_attn_func,
                     heads_k_stride=cfg.heads_k_stride,
-                    gather_outputs=cfg.rl is RLType.GRPO,
                 )
             )
 
@@ -545,6 +545,8 @@ def train(
     Returns:
         Tuple of (model, tokenizer) after training
     """
+    print_axolotl_text_art()
+
     # Setup model, tokenizer, (causal or RLHF) trainer, etc.
     (
         trainer,

src/axolotl/utils/config/__init__.py

@@ -21,7 +21,7 @@ from axolotl.utils.schemas.config import (
 from axolotl.utils.schemas.config import AxolotlInputConfig as AxolotlInputConfigBase
 from axolotl.utils.schemas.datasets import DPODataset, KTODataset, SFTDataset
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 
 def choose_device(cfg):

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -174,8 +174,6 @@ class SequenceParallelContextManager:
         ring_attn_func: Which ring attention function to use. Currently unused.
         heads_k_stride: Sequence parallelism K head stride size. Passed through to
             `varlen_llama3` `ring_flash_attn` implementation.
-        gather_outputs: Whether to gather outputs after model forward pass across the
-            sequence parallel group.
     """
 
     def __init__(
@@ -185,15 +183,12 @@ class SequenceParallelContextManager:
         gradient_accumulation_steps: int,
         ring_attn_func: RingAttnFunc,
         heads_k_stride: int | None,
-        gather_outputs: bool,
     ):
         self.models = models
         self.sequence_parallel_degree = sequence_parallel_degree
         self.gradient_accumulation_steps = gradient_accumulation_steps
         self.ring_attn_func = ring_attn_func
         self.heads_k_stride = heads_k_stride
-        self.gather_outputs = gather_outputs
-
         self._register_ring_attn()
 
         # Set distributed info for local rank
@@ -282,17 +277,16 @@ class SequenceParallelContextManager:
 
             return output
 
-        # Register hooks
+        # Register both hooks
         for model in self.models:
             self.hook_handles.append(
                 model.register_forward_pre_hook(
                     sequence_parallel_pre_hook, with_kwargs=True
                 )
             )
-            if self.gather_outputs:
+            self.hook_handles.append(
-                self.hook_handles.append(
+                model.register_forward_hook(sequence_parallel_post_hook)
-                    model.register_forward_hook(sequence_parallel_post_hook)
+            )
-                )
 
     def _gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
         """Gather sharded outputs from all ranks and reconstruct the full tensor."""

src/axolotl/utils/distributed.py

@@ -1,4 +1,6 @@
-"""Utilities for distributed functionality."""
+"""
+utility helpers for distributed checks
+"""
 
 import os
 import pickle  # nosec
@@ -17,7 +19,7 @@ from transformers.utils.import_utils import (
 distributed_state = None  # pylint: disable=invalid-name
 
 
-def get_device_type() -> torch.device:
+def get_device_type():
     device = torch.device("cpu")
     if is_torch_cuda_available():
         device = torch.device("cuda")
@@ -28,7 +30,7 @@ def get_device_type() -> torch.device:
     return device
 
 
-def get_device_count() -> int:
+def get_device_count():
     cur_device = get_device_type()
     if "cuda" in str(cur_device):
         return torch.cuda.device_count()
@@ -37,7 +39,7 @@ def get_device_count() -> int:
     return 1
 
 
-def get_current_device() -> int:
+def get_current_device():
     cur_device = get_device_type()
     if "cuda" in str(cur_device):
         return torch.cuda.current_device()
@@ -46,24 +48,15 @@ def get_current_device() -> int:
     return 0
 
 
-def init_distributed_state():
+def is_distributed():
+    """
+    Check if distributed training is initialized.
+    """
     global distributed_state  # pylint: disable=global-statement
-    if distributed_state is None:
+    if not distributed_state:
         timeout = int(os.environ.get("AXOLOTL_NCCL_TIMEOUT", 1800))
         distributed_state = PartialState(timeout=timedelta(seconds=timeout))
 
-
-def get_distributed_state() -> PartialState | None:
-    return distributed_state
-
-
-def is_distributed() -> bool:
-    """Check if distributed training is initialized."""
-    init_distributed_state()
-
-    if distributed_state is None:
-        return False
-
     return distributed_state.use_distributed and distributed_state.initialized
 
 
@@ -76,31 +69,31 @@ def barrier():
         dist.barrier()
 
 
-def is_main_process() -> bool:
+def is_main_process(use_environ=False):
     """
     Check if the current process is the main process. If not in distributed mode,
     always return `True`.
 
-    We use a simpler logic when the distributed state is not initialized: we just log
+    Args:
-    on the 0-th local rank.
+    - use_environ (bool, optional): Use environment variable to determine main process.
 
     Returns:
-        `True` if the current process is the main process, `False` otherwise.
+    - bool: `True` if the current process is the main process, `False` otherwise.
     """
-    if get_distributed_state() is None:
+    if use_environ:
         return os.environ.get("LOCAL_RANK", "0") == "0"
     if not is_distributed():
         return True
     return dist.get_rank() == 0
 
 
-def is_local_main_process() -> bool:
+def is_local_main_process(use_environ=False):
-    if get_distributed_state() is None:
+    if use_environ:
         return os.environ.get("LOCAL_RANK", "0") == "0"
     return PartialState().is_local_main_process
 
 
-def get_world_size() -> int:
+def get_world_size():
     return int(os.getenv("WORLD_SIZE", "1"))
 
 
@@ -122,7 +115,7 @@ def cleanup_distributed():
 
 
 @contextmanager
-def zero_first(is_main: bool):
+def zero_first(is_main):
     """
     runs the wrapped context so that rank 0 runs first before other ranks
     """

src/axolotl/utils/freeze.py

@@ -5,8 +5,9 @@ module to freeze/unfreeze parameters by name
 import re
 from typing import Callable, List, Tuple, Union
 
+from accelerate.logging import get_logger
+
 from axolotl.utils.distributed import is_main_process
-from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 

src/axolotl/utils/logging.py

@@ -1,4 +1,6 @@
-"""Logging helpers to only log on main process."""
+"""
+logging helpers to only log on main process
+"""
 
 import functools
 import logging
@@ -12,18 +14,27 @@ from axolotl.utils.distributed import is_main_process
 
 class MultiProcessAdapter(logging.LoggerAdapter):
     """
-    Logger adapter for distributed logging, specifically to only log on main process.
+    logger adapter for distributed logging, specifically to only log on main process
     """
 
+    def __init__(self, logger, use_environ=False, extra=None):
+        super().__init__(logger, extra)
+        self.use_environ = use_environ
+
     @staticmethod
-    def _should_log(main_process_only: bool):
+    def _should_log(main_process_only, use_environ=False):
-        return not main_process_only or is_main_process()
+        return not main_process_only or (
+            main_process_only and is_main_process(use_environ=use_environ)
+        )
 
     def log(self, level, msg, *args, **kwargs):
+        use_environ = kwargs.pop("use_environ", self.use_environ)
         main_process_only = kwargs.pop("main_process_only", True)
         kwargs.setdefault("stacklevel", 2)
 
-        if self.isEnabledFor(level) and self._should_log(main_process_only):
+        if self.isEnabledFor(level) and self._should_log(
+            main_process_only, use_environ=use_environ
+        ):
             msg, kwargs = self.process(msg, kwargs)
             self.logger.log(level, msg, *args, **kwargs)
 
@@ -39,11 +50,13 @@ class MultiProcessAdapter(logging.LoggerAdapter):
         self.warning(*args, **kwargs)
 
 
-def get_logger(name: str, log_level: str | None = None) -> MultiProcessAdapter:
+def get_logger(
+    name: str, log_level: str | None = None, use_environ: bool = False
+) -> MultiProcessAdapter:
     if log_level is None:
         log_level = os.environ.get("AXOLOTL_LOG_LEVEL", None)
     logger = logging.getLogger(name)
     if log_level is not None:
         logger.setLevel(log_level.upper())
         logger.root.setLevel(log_level.upper())
-    return MultiProcessAdapter(logger, extra={})
+    return MultiProcessAdapter(logger, use_environ=use_environ, extra={})

src/axolotl/utils/samplers/multipack.py

@@ -127,7 +127,7 @@ def pack_parallel(
     bin_size: int,
     num_processes: int | None = None,
     safe_mode: bool = True,
-    mp_start_method: str | None = "fork",
+    mp_start_method: str | None = "spawn",
 ) -> list[list[int]]:
     """Pack sequences into bins using parallel processing.
 
@@ -266,7 +266,6 @@ class MultipackBatchSampler(BatchSampler):
         bin_size: int = 200,  # The max number of samples that can be packed in a single bin
         num_processes: int | None = None,  # Number of processes for parallel packing
         safe_mode: bool = True,  # Conservative packing to prevent training instability
-        mp_start_method: str = "fork",
         **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
@@ -279,7 +278,6 @@ class MultipackBatchSampler(BatchSampler):
         self.bin_size = bin_size
         self.num_processes = num_processes
         self.safe_mode = safe_mode
-        self.mp_start_method = mp_start_method
 
         assert isinstance(self.lengths, np.ndarray)
 
@@ -340,9 +338,8 @@ class MultipackBatchSampler(BatchSampler):
                 bin_capacity=self.batch_max_len,
                 group_size=self.group_size,
                 bin_size=self.bin_size,
-                num_processes=max(4, self.num_processes) if self.num_processes else 4,
+                num_processes=self.num_processes,
                 safe_mode=self.safe_mode,
-                mp_start_method=self.mp_start_method,
             )
 
             # Map bin indices back to original indices

src/axolotl/utils/schemas/config.py

@@ -48,7 +48,7 @@ from axolotl.utils.schemas.trl import TRLConfig
 from axolotl.utils.schemas.validation import ValidationMixin
 from axolotl.utils.schemas.vllm import VllmConfig
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 
 # pylint: disable=too-many-ancestors
@@ -393,12 +393,6 @@ class AxolotlInputConfig(
         default=None,
         json_schema_extra={"description": "Whether to pack samples sequentially"},
     )
-    sample_packing_mp_start_method: str | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or 'forkserver'"
-        },
-    )
     eval_sample_packing: bool | None = Field(
         default=None,
         json_schema_extra={
@@ -529,19 +523,6 @@ class AxolotlInputConfig(
         },
     )
 
-    chunked_cross_entropy: bool | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Whether to use chunked cross entropy loss for memory efficiency"
-        },
-    )
-    chunked_cross_entropy_num_chunks: int | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "Number of chunks to use for chunked cross entropy loss"
-        },
-    )
-
     llama4_linearized_experts: bool | None = None
 
     deepspeed: str | dict[str, Any] | None = Field(

src/axolotl/utils/schemas/enums.py

@@ -54,7 +54,6 @@ class ChatTemplate(str, Enum):
     jinja = "jinja"
     qwen_25 = "qwen_25"
     qwen3 = "qwen3"
-    falcon_h1 = "falcon_h1"
     tokenizer_default = "tokenizer_default"
     exaone = "exaone"
     metharme = "metharme"

src/axolotl/utils/schemas/model.py

@@ -4,7 +4,7 @@ from pydantic import BaseModel, Field, field_validator
 
 from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = get_logger(__name__, use_environ=True)
 
 
 class ModelInputConfig(BaseModel):

src/axolotl/utils/trainer.py

@@ -11,14 +11,14 @@ from typing import List, Optional
 import numpy as np
 import torch
 import torch.cuda
+from accelerate.logging import get_logger
 from datasets import IterableDataset, disable_caching, enable_caching
 from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.monkeypatch.trainer_eval_guard import patch_evaluation_loop_for_fsdp2
-from axolotl.utils.distributed import init_distributed_state, reduce_and_broadcast
+from axolotl.utils.distributed import reduce_and_broadcast
 from axolotl.utils.environment import check_cuda_p2p_ib_support
-from axolotl.utils.logging import get_logger
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger(__name__)
@@ -467,7 +467,6 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
                 sequential=cfg.sample_packing_sequentially,
                 drop_last=True,
                 num_processes=cfg.dataset_processes,
-                mp_start_method=cfg.sample_packing_mp_start_method or "fork",
             )
 
             data_loader = DataLoader(
@@ -538,12 +537,6 @@ def setup_deepspeed_env(cfg, stage=None):
         os.environ["ACCELERATE_DEEPSPEED_ZERO_STAGE"] = str(stage)
         if stage == 3:
             os.environ["ACCELERATE_DEEPSPEED_ZERO3_INIT"] = "true"
-
-    # NOTE(djsaunde): The distribued state cannot be initialized prior to the
-    # ACCELERATE_USE_DEEPSPEED assignment, but it must be initialized some time prior
-    # to model load.
-    init_distributed_state()
-
     # If we don't assign this, it doesn't actually get set in the accelerate weakref
     _ = HfTrainerDeepSpeedConfig(cfg.deepspeed)
 

tests/test_chunked_xentropy.py

@@ -1,40 +0,0 @@
-"""
-test suite for chunked cross entropy
-"""
-
-import pytest
-import torch
-from torch import nn
-
-from axolotl.monkeypatch.loss.chunked import get_causal_lm_loss
-
-
-@pytest.fixture
-def chunked_fixtures():
-    model_dim = 512
-    vocab_size = 1024 * 256
-    seq_len = 2048
-    batch_size = 1
-
-    lm_head = nn.Linear(model_dim, vocab_size)
-    hidden_state = torch.randn(batch_size, seq_len, model_dim)
-    labels = torch.randint(low=0, high=vocab_size, size=(batch_size, seq_len))
-    return lm_head, hidden_state, labels, vocab_size
-
-
-def test_chunked_forward(chunked_fixtures):  # pylint: disable=redefined-outer-name
-    lm_head, hidden_state, labels, vocab_size = chunked_fixtures
-    lm_loss = get_causal_lm_loss()
-
-    logits = lm_head(hidden_state)
-
-    chunked_lm_loss = lm_loss(logits, labels)
-
-    logits_flattened = logits.view(-1, vocab_size)
-    labels_flattened = labels.view(-1)
-
-    loss = nn.functional.cross_entropy(
-        logits_flattened.float(), labels_flattened, reduction="mean"
-    )
-
-    assert torch.allclose(chunked_lm_loss, loss, atol=1e-2, rtol=1e-2)