fix for gather across multiple gpus

fix: finetune model inference needs the dtype fix to work with flash-attn

.github/workflows/main.yml

@@ -13,17 +13,17 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: cu118
+          - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.9"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: cu118
+          - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: cu118
+          - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.9"
             pytorch: 2.0.1
@@ -49,10 +49,11 @@ jobs:
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-${{ matrix.cuda }}-${{ matrix.pytorch }}
+            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
-          tags: ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          tags: ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}
   build-axolotl-runpod:
     needs: build-axolotl

.github/workflows/tests.yml

@@ -24,7 +24,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip install -e .
+          pip install -e .[peft]
           pip install -r requirements-tests.txt
 
       - name: Run tests

README.md

@@ -16,6 +16,7 @@ Axolotl is a tool designed to streamline the fine-tuning of various AI models, o
   - [LambdaLabs Installation](#lambdalabs)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
+  - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
 - [Config](#config)
   - [Train](#train)
   - [Inference](#inference)
@@ -68,8 +69,9 @@ Get started with Axolotl in just a few steps! This quickstart guide will walk yo
 
 ```bash
 git clone https://github.com/OpenAccess-AI-Collective/axolotl
+cd axolotl
 
-pip3 install -e .
+pip3 install -e .[flash-attn]
 pip3 install -U git+https://github.com/huggingface/peft.git
 
 # finetune lora
@@ -98,7 +100,7 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
   ```
 
 - Conda/Pip venv
-  1. Install python **3.9**
+  1. Install python >=**3.9**
 
   2. Install pytorch stable https://pytorch.org/get-started/locally/
 
@@ -151,9 +153,7 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
 
   pip3 install -e . # change depend on needs
   pip3 install protobuf==3.20.3
-  pip3 install -U requests
-  pip3 install -U --ignore-installed psutil
-  pip3 install -U scipy
+  pip3 install -U --ignore-installed requests Pillow psutil scipy
   pip3 install git+https://github.com/huggingface/peft.git # not for gptq
   ```
 
@@ -257,6 +257,10 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"conversations": [{"role": "...", "value": "..."}]}
   ```
+- `metharme`: instruction, adds additional eos tokens
+  ```json
+  {"prompt": "...", "generation": "..."}
+  ```
 - `sharegpt_simple.load_role`: conversations where `role` is used instead of `from`
   ```json
   {"conversations": [{"role": "...", "value": "..."}]}
@@ -274,11 +278,29 @@ Have dataset(s) in one of the following format (JSONL recommended):
 
 #### How to add custom prompts
 
-  1. Add your method to a file in [prompt_strategies](src/axolotl/prompt_strategies). Please see other files as example.
-  2. Use your custom file name as the dataset type `<prompt_strategies_file>.load_<load_fn>`.
+Using yaml. Example:
+```yaml
+datasets:
+  - path: repo
+    type:
+      system_prompt: ""
+      no_input_format: |-
+        User: {instruction}<|end_of_turn|>
+        Assistant:
+      format: |-
+        User: {instruction}
+        {input}<|end_of_turn|>
+        Assistant:
+```
 
-Optionally, download some datasets, see [data/README.md](data/README.md)
+Using file:
+1. Add your method to a file in [prompt_strategies](src/axolotl/prompt_strategies). Please see other files as example.
+2. Use your custom file name as the dataset type `<prompt_strategies_file>.load_<load_fn>`.
 
+#### How to use your custom pretokenized dataset
+
+- Do not pass a `type:`
+- Dataset must contain `input_ids`, `attention_mask`, `labels` in columns
 
 
 ### Config
@@ -308,9 +330,9 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 
   # local
   datasets:
-    - path: json
-      data_files: data.jsonl # or json
-      type: alpaca # format from earlier
+    - path: data.jsonl # or json
+      ds_type: json # see other options below
+      type: alpaca
   ```
 
 - loading
@@ -391,10 +413,29 @@ datasets:
   - path: vicgalle/alpaca-gpt4
   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
+    ds_type: # Optional[str] (json|arrow|parquet) defines the datatype when path is a file
     data_files: # path to source data files
     shards: # number of shards to split data into
     name: # name of dataset configuration to load
 
+  # custom user prompt
+  - path: repo
+    type:
+      # the below are defaults. only set what's needed.
+      system_prompt: ""
+      field_system: system
+      field_instruction: instruction
+      field_output: input
+
+      # customizable to be single line or multi-line
+      system_format: "{system}"
+      # 'format' can include {input}
+      format: |-
+        User: {instruction} {input}
+        Assistant:
+      # 'no_input_format' cannot include {input}
+      no_input_format: "{instruction} "
+
 # axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
@@ -452,6 +493,12 @@ lora_modules_to_save:
 lora_out_dir:
 lora_fan_in_fan_out: false
 
+# ReLoRA configuration
+# must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+relora_steps: # number of steps per ReLoRA restart
+relora_warmup_steps: # number of per-restart warmup steps
+relora_cpu_offload: # true to perform lora weight merges on cpu during restarts, for modest gpu memory savings
+
 # wandb configuration if you're using it
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
 wandb_project: # your wandb project name
@@ -472,8 +519,9 @@ warmup_steps: 100
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_steps:
+save_strategy: # set to `no` to skip checkpoint saves
 save_steps: # leave empty to save at each epoch
-eval_steps:
+eval_steps: # leave empty to eval at each epoch
 save_total_limit: # checkpoints saved at a time
 max_steps:
 
@@ -584,7 +632,7 @@ strict:
 
 Run
 ```bash
-accelerate launch scripts/finetune.py configs/your_config.yml
+accelerate launch scripts/finetune.py your_config.yml
 ```
 
 #### Multi-GPU
@@ -666,7 +714,9 @@ Please reduce any below
   - `gradient_accumulation_steps`
   - `sequence_len`
 
-> `failed (exitcode: -9)` usually means your system has run out of system memory.
+> `failed (exitcode: -9)`
+
+Usually means your system has run out of system memory.
 Similarly, you should consider reducing the same settings as when you run out of VRAM.
 Additionally, look into upgrading your system RAM which should be simpler than GPU upgrades.
 

data/README.md

@@ -1,24 +0,0 @@
-
-## Download some datasets
-```shell
-curl https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_gpt4.json -o data/raw/alpaca_data_gpt4.json
-curl https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json -L -o data/raw/vicuna_cleaned.json
-curl https://github.com/teknium1/GPTeacher/blob/main/Instruct/gpt4-instruct-similarity-0.6-dataset.json?raw=true -L -o data/raw/gpt4-instruct-similarity-0.6-dataset.json
-curl https://github.com/teknium1/GPTeacher/blob/main/Roleplay/roleplay-similarity_0.6-instruct-dataset.json?raw=true -L -o data/raw/roleplay-similarity_0.6-instruct-dataset.json
-```
-
-## Convert the JSON data files to JSONL.
-
-```shell
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/alpaca_data_gpt4.json --output data/alpaca_data_gpt4.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/vicuna_cleaned.json --output data/vicuna_cleaned.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/roleplay-similarity_0.6-instruct-dataset.json --output data/roleplay-similarity_0.6-instruct-dataset.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/gpt4-instruct-similarity-0.6-dataset.json --output data/gpt4-instruct-similarity-0.6-dataset.jsonl
-```
----
-
-Using JSONL makes it easier to subset the data if you want a smaller training set, i.e get 2000 random examples.
-
-```shell
-shuf -n2000 data/vicuna_cleaned.jsonl > data/vicuna_cleaned.subset0.jsonl
-```

data/raw/.gitignore

@@ -1 +0,0 @@
-**

deepspeed/zero2.json

@@ -0,0 +1,46 @@
+{
+    "zero_optimization": {
+      "stage": 2,
+      "offload_optimizer": {
+        "device": "cpu"
+      },
+      "contiguous_gradients": true,
+      "overlap_comm": true
+    },
+    "bf16": {
+      "enabled": "auto"
+    },
+    "fp16": {
+      "enabled": "auto",
+      "auto_cast": false,
+      "loss_scale": 0,
+      "initial_scale_power": 32,
+      "loss_scale_window": 1000,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+    },
+    "optimizer": {
+      "type": "AdamW",
+      "params": {
+        "lr": "auto",
+        "betas": [
+          0.9,
+          0.999
+        ],
+        "eps": 1e-8,
+        "weight_decay": "auto"
+      }
+    },
+    "scheduler": {
+      "type": "WarmupDecayLR",
+      "params": {
+        "warmup_min_lr": "auto",
+        "warmup_max_lr": "auto",
+        "warmup_num_steps": "auto",
+        "total_num_steps": "auto"
+      }
+    },
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

docker/Dockerfile

@@ -16,9 +16,9 @@ RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN cd axolotl && \
     if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[$AXOLOTL_EXTRAS]; \
+        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .; \
+        pip install -e .[flash-attn]; \
     fi
 
 # fix so that git fetch/pull from remote works

docker/Dockerfile-base

@@ -31,26 +31,6 @@ WORKDIR /workspace
 RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA
 
-
-FROM base-builder AS flash-attn-builder
-
-WORKDIR /workspace
-
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-
-RUN git clone https://github.com/Dao-AILab/flash-attention.git && \
-    cd flash-attention && \
-    git checkout v2.0.4  && \
-    python3 setup.py bdist_wheel && \
-    cd csrc/fused_dense_lib && \
-    python3 setup.py bdist_wheel && \
-    cd ../xentropy && \
-    python3 setup.py bdist_wheel && \
-    cd ../rotary && \
-    python3 setup.py bdist_wheel && \
-    cd ../layer_norm && \
-    python3 setup.py bdist_wheel
-
 FROM base-builder AS deepspeed-builder
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
@@ -90,13 +70,8 @@ RUN mkdir -p /workspace/wheels/bitsandbytes
 COPY --from=deepspeed-builder /workspace/DeepSpeed/dist/deepspeed-*.whl wheels
 COPY --from=bnb-builder /workspace/bitsandbytes/dist/bitsandbytes-*.whl wheels
 COPY --from=bnb-builder /workspace/bitsandbytes/bitsandbytes/libbitsandbytes*.so wheels/bitsandbytes
-COPY --from=flash-attn-builder /workspace/flash-attention/dist/flash_attn-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/fused_dense_lib/dist/fused_dense_lib-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/xentropy/dist/xentropy_cuda_lib-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/rotary/dist/rotary_emb-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/layer_norm/dist/dropout_layer_norm-*.whl wheels
 
-RUN pip3 install wheels/deepspeed-*.whl wheels/flash_attn-*.whl wheels/fused_dense_lib-*.whl wheels/xentropy_cuda_lib-*.whl wheels/rotary_emb-*.whl wheels/dropout_layer_norm-*.whl
+RUN pip3 install wheels/deepspeed-*.whl
 RUN cd /workspace/builds/bitsandbytes && python3 setup.py install
 RUN git lfs install --skip-repo
 RUN pip3 install awscli && \

examples/code-llama/13b/lora.yml

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/13b/qlora.yml

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/34b/lora.yml

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/34b/qlora.yml

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/7b/lora.yml

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/7b/qlora.yml

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/README.md

@@ -0,0 +1,22 @@
+# Overview
+
+This is an example of CodeLLaMA configuration for 7b, 13b and 34b.
+
+The 7b variant fits on any 24GB VRAM GPU and will take up about 17 GB of VRAM during training if using qlora and 20 GB if using lora. On a RTX 4090 it trains 3 epochs of the default dataset in about 15 minutes.
+
+The 13b variant will fit if you change these settings to these values:
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+
+The 34b variant does not fit on 24GB of VRAM - you will need something with +40 gb VRAM that also supports flash attention v2 - A6000 or A100 are good choices.
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/qlora.yml
+
+```
+or
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/lora.yml
+
+```

examples/gptq-lora-7b/config.yml

@@ -57,7 +57,7 @@ weight_decay: 0.0001
 fsdp:
 fsdp_config:
 tokens:
-  pad_token: "[PAD]"
+  pad_token: "<pad>"
   bos_token: "<s>"
   eos_token: "</s>"
   unk_token: "<unk>"

examples/llama-2/relora.yml

@@ -0,0 +1,73 @@
+base_model: meta-llama/Llama-2-7b-hf
+base_model_config: meta-llama/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./relora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+relora_steps: 150
+relora_warmup_steps: 10
+relora_cpu_offload: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+save_steps: 50
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

requirements.txt

@@ -1,20 +1,23 @@
+packaging
 peft @ git+https://github.com/huggingface/peft.git
 transformers @ git+https://github.com/huggingface/transformers.git
 bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@2a289f6108e77a77a4efffb3f6316bc98538413b
 addict
+evaluate
 fire
-PyYAML==6.0
+PyYAML>=6.0
 datasets
-accelerate>=0.19.0
+flash-attn>=2.0.8
 sentencepiece
 wandb
 einops
 xformers
 optimum
 hf_transfer
+colorama
 numba
-numpy==1.24.4
+numpy>=1.24.4
 # qlora things
 bert-score==0.3.13
 evaluate==0.4.0

scripts/alpaca_json_to_jsonl.py

@@ -1,52 +0,0 @@
-"""Module to convert json file to jsonl"""
-
-import os
-import sys
-from pathlib import Path
-from typing import Optional, Union
-
-import fire
-
-from axolotl.convert import (
-    FileReader,
-    FileWriter,
-    JsonlSerializer,
-    JsonParser,
-    JsonToJsonlConverter,
-    StdoutWriter,
-)
-from axolotl.logging_config import configure_logging
-
-configure_logging()
-
-# add src to the pythonpath so we don't need to pip install this
-project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
-src_dir = os.path.join(project_root, "src")
-sys.path.insert(0, src_dir)
-
-
-def main(
-    file: Path,
-    output: Optional[Path] = None,
-    to_stdout: Optional[bool] = False,
-):
-    """
-    Convert a json file to jsonl
-    """
-
-    file_reader = FileReader()
-    writer: Union[StdoutWriter, FileWriter]
-    if to_stdout or output is None:
-        writer = StdoutWriter()
-    else:
-        writer = FileWriter(output)
-    json_parser = JsonParser()
-    jsonl_serializer = JsonlSerializer()
-
-    converter = JsonToJsonlConverter(file_reader, writer, json_parser, jsonl_serializer)
-
-    converter.convert(file, output)
-
-
-if __name__ == "__main__":
-    fire.Fire(main)

scripts/finetune.py

@@ -82,6 +82,8 @@ def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
             max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
         )
 
+    model = model.to(cfg.device)
+
     while True:
         print("=" * 80)
         # support for multiline inputs
@@ -242,6 +244,21 @@ def train(
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
         return
 
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [
+            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
+        ]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(
+                possible_checkpoints,
+                key=lambda path: int(path.split("-")[-1]),
+            )
+            cfg.resume_from_checkpoint = sorted_paths[-1]
+            LOG.info(
+                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
+            )
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+
     trainer = setup_trainer(
         cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
     )
@@ -273,20 +290,6 @@ def train(
     LOG.info("Starting trainer...")
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
-    resume_from_checkpoint = cfg.resume_from_checkpoint
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
-        possible_checkpoints = [
-            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
-        ]
-        if len(possible_checkpoints) > 0:
-            sorted_paths = sorted(
-                possible_checkpoints,
-                key=lambda path: int(path.split("-")[-1]),
-            )
-            resume_from_checkpoint = sorted_paths[-1]
-            LOG.info(
-                f"Using Auto-resume functionality to start with checkpoint at {resume_from_checkpoint}"
-            )
 
     if not Path(cfg.output_dir).is_dir():
         os.makedirs(cfg.output_dir, exist_ok=True)
@@ -301,6 +304,13 @@ def train(
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
+    if cfg.relora_steps:
+        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
+            model = model.merge_and_unload()
+        else:
+            # final model weights have already been saved by `ReLoRACallback.on_train_end`
+            return
+
     # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
     # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
     if cfg.fsdp:
@@ -308,6 +318,7 @@ def train(
     elif cfg.local_rank == 0:
         if cfg.flash_optimum:
             model = BetterTransformer.reverse(model)
+
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
 

setup.py

@@ -7,6 +7,7 @@ with open("./requirements.txt", encoding="utf-8") as requirements_file:
     # don't include peft yet until we check the int4
     # need to manually install peft for now...
     reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
+    reqs = [r for r in reqs if "flash-attn" not in r]
     reqs = [r for r in reqs if r and r[0] != "#"]
     for r in reqs:
         install_requires.append(r)
@@ -25,9 +26,14 @@ setup(
         "gptq_triton": [
             "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
         ],
+        "flash-attn": [
+            "flash-attn==2.0.8",
+        ],
         "extras": [
-            "flash-attn",
             "deepspeed",
         ],
+        "peft": [
+            "peft @ git+https://github.com/huggingface/peft.git",
+        ],
     },
 )

src/axolotl/logging_config.py

@@ -1,16 +1,42 @@
-"""Logging configuration settings"""
+"""
+Common logging module for axolotl
+"""
 
 import os
 import sys
+from logging import Formatter
 from logging.config import dictConfig
 from typing import Any, Dict
 
+from colorama import Fore, Style, init
+
+
+class ColorfulFormatter(Formatter):
+    """
+    Formatter to add coloring to log messages by log type
+    """
+
+    COLORS = {
+        "WARNING": Fore.YELLOW,
+        "ERROR": Fore.RED,
+        "CRITICAL": Fore.RED + Style.BRIGHT,
+    }
+
+    def format(self, record):
+        log_message = super().format(record)
+        return self.COLORS.get(record.levelname, "") + log_message + Fore.RESET
+
+
 DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
     "version": 1,
     "formatters": {
         "simple": {
             "format": "[%(asctime)s] [%(levelname)s] [%(name)s.%(funcName)s:%(lineno)d] [PID:%(process)d] %(message)s",
         },
+        "colorful": {
+            "()": ColorfulFormatter,
+            "format": "[%(asctime)s] [%(levelname)s] [%(name)s.%(funcName)s:%(lineno)d] [PID:%(process)d] %(message)s",
+        },
     },
     "filters": {},
     "handlers": {
@@ -20,14 +46,25 @@ DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
             "filters": [],
             "stream": sys.stdout,
         },
+        "color_console": {
+            "class": "logging.StreamHandler",
+            "formatter": "colorful",
+            "filters": [],
+            "stream": sys.stdout,
+        },
     },
     "root": {"handlers": ["console"], "level": os.getenv("LOG_LEVEL", "INFO")},
     "loggers": {
-        "axolotl": {"handlers": ["console"], "level": "DEBUG", "propagate": False},
+        "axolotl": {
+            "handlers": ["color_console"],
+            "level": "DEBUG",
+            "propagate": False,
+        },
     },
 }
 
 
 def configure_logging():
     """Configure with default logging"""
+    init()  # Initialize colorama
     dictConfig(DEFAULT_LOGGING_CONFIG)

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,142 +2,47 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
-from typing import Optional, Tuple
+import warnings
+from typing import List, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer as OriginalLlamaDecoderLayer,
+)
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
 try:
-    from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
+    from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
+        flash_attn_kvpacked_func,
+        flash_attn_varlen_kvpacked_func,
+        flash_attn_varlen_qkvpacked_func,
+    )
 except ImportError:
+    from flash_attn.flash_attn_interface import (
+        flash_attn_unpadded_kvpacked_func as flash_attn_varlen_kvpacked_func,
+    )
     from flash_attn.flash_attn_interface import (
         flash_attn_unpadded_qkvpacked_func as flash_attn_varlen_qkvpacked_func,
     )
 
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
 
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-
-
-def forward(
-    self,
-    hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.Tensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
-    output_attentions: bool = False,
-    use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    """Input shape: Batch x Time x Channel
-
-    attention_mask: [bsz, q_len]
-    """
-    # pylint: disable=duplicate-code
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
+def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
+        _prepare_decoder_attention_mask
     )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    # [bsz, q_len, nh, hd]
-    # [bsz, nh, q_len, hd]
-
-    kv_seq_len = key_states.shape[-2]
-    assert past_key_value is None, "past_key_value is not supported"
-
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    query_states, key_states = apply_rotary_pos_emb(
-        query_states, key_states, cos, sin, position_ids
-    )
-    # [bsz, nh, t, hd]
-    assert not output_attentions, "output_attentions is not supported"
-    assert not use_cache, "use_cache is not supported"
-
-    # Flash attention codes from
-    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
-
-    # transform the data into the format required by flash attention
-    qkv = torch.stack(
-        [query_states, key_states, value_states], dim=2
-    )  # [bsz, nh, 3, q_len, hd]
-    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
-    # We have disabled _prepare_decoder_attention_mask in LlamaModel
-    # the attention_mask should be the same as the key_padding_mask
-    key_padding_mask = attention_mask
-
-    if key_padding_mask is None:
-        qkv = rearrange(qkv, "b s ... -> (b s) ...")
-        max_s = q_len
-        cu_q_lens = torch.arange(
-            0,
-            (bsz + 1) * q_len,
-            step=q_len,
-            dtype=torch.int32,
-            device=qkv.device,
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = flashattn_forward
+    if packed:
+        transformers.models.llama.modeling_llama.LlamaDecoderLayer = LlamaDecoderLayer
+        transformers.models.llama.modeling_llama.LlamaModel.forward = (
+            llama_model_forward
         )
-        output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
-        )
-        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
-    elif attention_mask.shape[0] == 1:
-        # special handling using sample packing
-        qkv = rearrange(qkv, "b s ... -> (b s) ...")
-        cu_q_lens, max_s = get_cu_seqlens_from_pos_ids(position_ids)
-        cu_q_lens = cu_q_lens.squeeze()
-
-        output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
-        )
-        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
-    else:
-        nheads = qkv.shape[-2]
-
-        # pylint: disable=invalid-name
-        x = rearrange(qkv, "b s three h d -> b s (three h d)")
-        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
-        x_unpad = rearrange(
-            x_unpad,
-            "nnz (three h d) -> nnz three h d",
-            three=3,
-            h=nheads,
-        )
-        output_unpad = flash_attn_varlen_qkvpacked_func(
-            x_unpad,
-            cu_q_lens,
-            max_s,
-            0.0,
-            softmax_scale=None,
-            causal=True,
-        )
-        output = rearrange(
-            pad_input(
-                rearrange(output_unpad, "nnz h d -> nnz (h d)"),
-                indices,
-                bsz,
-                q_len,
-            ),
-            "b s (h d) -> b s h d",
-            h=nheads,
-        )
-
-    return (
-        self.o_proj(rearrange(output, "b s h d -> b s (h d)")),
-        None,
-        None,
-    )
 
 
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
@@ -153,8 +58,541 @@ def _prepare_decoder_attention_mask(
     return attention_mask
 
 
-def replace_llama_attn_with_flash_attn():
-    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
-        _prepare_decoder_attention_mask
+def flashattn_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.Tensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    """Input shape: Batch x Time x Channel
+
+    attention_mask: [bsz, q_len]
+    """
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        value_states = torch.cat(value_states, dim=-1)
+
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+    query_states = query_states.view(
+        bsz, q_len, self.num_heads, self.head_dim
+    ).transpose(1, 2)
+    key_states = key_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    value_states = value_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
     )
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
+    # [bsz, nh, t, hd]
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # flash-attn v2 start
+    #
+
+    if self.training:
+        # during training q,k,v always have same seqlen
+        assert key_states.shape == query_states.shape
+        is_causal = True
+    else:
+        # turn off FA causal mask after first inference autoregressive iteration
+        # only on first autoregressive step q,k,v have same seqlen
+        is_causal = key_states.shape == query_states.shape
+
+    if cu_seqlens is not None and max_seqlen is not None:
+        # special handling using sample packing
+        qkv = torch.stack(
+            [query_states, key_states, value_states], dim=2
+        )  # [bsz, nh, 3, q_len, hd]
+        qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+        qkv = rearrange(qkv, "b s ... -> (b s) ...")
+
+        output = flash_attn_varlen_qkvpacked_func(
+            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=None, causal=True
+        )
+        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
+    elif query_states.shape == key_states.shape:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        qkv_unpad, cu_seqlens_q, max_seqlen_q, _, output_pad_fn = generate_qkv(
+            query_states,
+            key_states,
+            value_states,
+            qkvpacked=True,
+            # We have disabled _prepare_decoder_attention_mask in LlamaModel
+            # the attention_mask should be the same as the key_padding_mask
+            key_padding_mask=attention_mask,
+            query_padding_mask=attention_mask[:, -query_states.size(1) :]
+            if attention_mask is not None
+            else None,
+        )
+        output_unpad = flash_attn_varlen_qkvpacked_func(
+            qkv_unpad,
+            cu_seqlens_q,
+            max_seqlen_q,
+            0.0,
+            softmax_scale=None,
+            causal=is_causal,
+        )
+        output = output_pad_fn(output_unpad)
+    else:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        if attention_mask is None or attention_mask.all().item():
+            output = flash_attn_kvpacked_func(
+                query_states,
+                torch.stack([key_states, value_states], 2),
+                causal=is_causal,
+            )
+        else:
+            (  # pylint: disable=unbalanced-tuple-unpacking
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                _,
+                _,
+                output_pad_fn,
+            ) = generate_qkv(
+                query_states,
+                key_states,
+                value_states,
+                kvpacked=True,
+                key_padding_mask=attention_mask,
+                query_padding_mask=attention_mask[:, -query_states.size(1) :]
+                if attention_mask is not None
+                else None,
+            )
+            output_unpad = flash_attn_varlen_kvpacked_func(
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                0.0,
+                softmax_scale=None,
+                causal=is_causal,
+            )
+            output = output_pad_fn(output_unpad)
+
+    attn_output = output
+    if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, q_len, self.num_heads, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+    attn_output = rearrange(attn_output, "b s h d -> b s (h d)")
+
+    #
+    # flash-attn v2 end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value
+
+
+# based on https://github.com/Dao-AILab/flash-attention/blob/364a5b/tests/test_flash_attn.py#L38
+def generate_qkv(
+    q,
+    k,
+    v,
+    query_padding_mask=None,
+    key_padding_mask=None,
+    kvpacked=False,
+    qkvpacked=False,
+):  # pylint: disable=invalid-name,unnecessary-lambda-assignment
+    """
+    Arguments:
+        q: (batch_size, seqlen_q, nheads, d)
+        k: (batch_size, seqlen_k, nheads_k, d)
+        v: (batch_size, seqlen_k, nheads_k, d)
+        query_padding_mask: (batch_size, seqlen), bool
+        key_padding_mask: (batch_size, seqlen), bool
+    """
+    assert not (kvpacked and qkvpacked)
+    batch_size, seqlen_q, nheads, d = q.shape
+    _, seqlen_k, nheads_k, _ = k.shape
+    assert k.shape == (batch_size, seqlen_k, nheads_k, d)
+    assert v.shape == (batch_size, seqlen_k, nheads_k, d)
+
+    if query_padding_mask is not None:
+        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(
+            q, query_padding_mask
+        )
+
+        output_pad_fn = lambda output_unpad: pad_input(  # noqa: E731
+            output_unpad, indices_q, batch_size, seqlen_q
+        )
+
+    else:
+        q_unpad = rearrange(q, "b s h d -> (b s) h d")
+        cu_seqlens_q = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_q,
+            step=seqlen_q,
+            dtype=torch.int32,
+            device=q_unpad.device,
+        )
+        max_seqlen_q = seqlen_q
+
+        output_pad_fn = lambda output_unpad: rearrange(  # noqa: E731
+            output_unpad, "(b s) h d -> b s h d", b=batch_size
+        )
+
+    if key_padding_mask is not None:
+        k_unpad, _, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
+        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
+    else:
+        k_unpad = rearrange(k, "b s h d -> (b s) h d")
+        v_unpad = rearrange(v, "b s h d -> (b s) h d")
+        cu_seqlens_k = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_k,
+            step=seqlen_k,
+            dtype=torch.int32,
+            device=k_unpad.device,
+        )
+        max_seqlen_k = seqlen_k
+
+    if qkvpacked:
+        assert nheads == nheads_k
+        qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
+        qkv = torch.stack([q, k, v], dim=2)
+        return (qkv_unpad, cu_seqlens_q, max_seqlen_q, qkv, output_pad_fn)
+
+    if kvpacked:
+        kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
+        kv = torch.stack([k, v], dim=2)
+        return (
+            q_unpad,
+            kv_unpad,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            q,
+            kv,
+            output_pad_fn,
+        )
+
+    return (
+        q_unpad,
+        k_unpad,
+        v_unpad,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        q,
+        k,
+        v,
+        output_pad_fn,
+    )
+
+
+def llama_model_forward(
+    self,
+    input_ids: 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,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    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
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+        )
+    if input_ids is not None:
+        batch_size, seq_length = input_ids.shape
+    elif inputs_embeds is not None:
+        batch_size, seq_length, _ = inputs_embeds.shape
+    else:
+        raise ValueError(
+            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+        )
+
+    seq_length_with_past = seq_length
+    past_key_values_length = 0
+
+    if past_key_values is not None:
+        past_key_values_length = past_key_values[0][0].shape[2]
+        seq_length_with_past = seq_length_with_past + past_key_values_length
+
+    cu_seqlens = None
+    max_seqlen = None
+    if position_ids is None:
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=device,
+        )
+        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+    else:
+        position_ids = position_ids.view(-1, seq_length).long()
+        cu_seqlens, max_seqlen = get_cu_seqlens_from_pos_ids(position_ids)
+        cu_seqlens = cu_seqlens.squeeze()
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+    # embed positions
+    if attention_mask is None:
+        attention_mask = torch.ones(
+            (batch_size, seq_length_with_past),
+            dtype=torch.bool,
+            device=inputs_embeds.device,
+        )
+    attention_mask = (
+        self._prepare_decoder_attention_mask(  # pylint: disable=protected-access
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+        )
+    )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            transformers.logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    next_decoder_cache = () if use_cache else None
+
+    for idx, decoder_layer in enumerate(self.layers):
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                position_ids,
+                None,
+                output_attentions,
+                None,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+            if v is not None
+        )
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )
+
+
+class LlamaDecoderLayer(OriginalLlamaDecoderLayer):
+    """
+    patched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens
+    """
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cu_seqlens (`torch.Tensor`, *optional*) cumulative sequence len when packing
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs

src/axolotl/monkeypatch/llama_attn_hijack_sdp.py

@@ -0,0 +1,140 @@
+"""
+Patched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention
+"""
+
+import warnings
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import transformers.models.llama.modeling_llama
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+
+def hijack_llama_sdp_attention():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
+        sdp_attention_forward
+    )
+
+
+def sdp_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        value_states = torch.cat(value_states, dim=-1)
+
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+    query_states = query_states.view(
+        bsz, q_len, self.num_heads, self.head_dim
+    ).transpose(1, 2)
+    key_states = key_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    value_states = value_states.view(
+        bsz, q_len, self.num_key_value_heads, self.head_dim
+    ).transpose(1, 2)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+    # [bsz, nh, t, hd]
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # repeat k/v heads if n_kv_heads < n_heads
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # sdp-attn start
+    #
+
+    with torch.backends.cuda.sdp_kernel():
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            is_causal=False,
+        )
+
+    if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+    attn_output = attn_output.transpose(1, 2)
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    #
+    # sdp-attn end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -3,13 +3,13 @@ Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-g
 """
 
 import logging
-import math
+import warnings
 from typing import Optional, Tuple
 
 import torch
 import torch.nn.functional as F
 import transformers.models.llama.modeling_llama
-from torch import nn
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
 try:
     import xformers.ops
@@ -21,12 +21,6 @@ def hijack_llama_attention():
     transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
 
 
-def hijack_llama_sdp_attention():
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
-        sdp_attention_forward
-    )
-
-
 def xformers_forward(
     self,
     hidden_states: torch.Tensor,
@@ -81,15 +75,15 @@ def xformers_forward(
     value_states = value_states.view(
         bsz, q_len, self.num_key_value_heads, self.head_dim
     ).transpose(1, 2)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
 
     kv_seq_len = key_states.shape[-2]
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
+
     cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    (
-        query_states,
-        key_states,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
+    query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )
     # [bsz, nh, t, hd]
@@ -102,74 +96,50 @@ def xformers_forward(
     past_key_value = (key_states, value_states) if use_cache else None
 
     # repeat k/v heads if n_kv_heads < n_heads
-    key_states = transformers.models.llama.modeling_llama.repeat_kv(
-        key_states, self.num_key_value_groups
-    )
-    value_states = transformers.models.llama.modeling_llama.repeat_kv(
-        value_states, self.num_key_value_groups
-    )
+    key_states = repeat_kv(key_states, self.num_key_value_groups)
+    value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-    # We only apply xformers optimizations if we don't need to output the whole attention matrix
-    if not output_attentions:
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
 
-        # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
-        # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
-        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states, key_states, value_states, attn_bias=None
-            )
-        else:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states,
-                key_states,
-                value_states,
-                # attn_bias=attention_mask,
-                attn_bias=xformers.ops.LowerTriangularMask(),
-            )
-        attn_weights = None
+    #
+    # xformers-attn start
+    #
+
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
+    # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
+    if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states, key_states, value_states, attn_bias=None
+        )
     else:
-        attn_weights = torch.matmul(
-            query_states, key_states.transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
-
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
-                f" {attn_weights.size()}"
-            )
-
-        if attention_mask is not None:
-            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
-                raise ValueError(
-                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
-                )
-            attn_weights = attn_weights + attention_mask
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(
-            attn_weights, dim=-1, dtype=torch.float32
-        ).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        # end x-formers vs. not x-formers if-else block
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states,
+            key_states,
+            value_states,
+            # attn_bias=attention_mask,
+            attn_bias=xformers.ops.LowerTriangularMask(),
+        )
 
+    if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, q_len, self.num_heads, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 
+    #
+    # xformers-attn end
+    #
+
     if self.pretraining_tp > 1:
         attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
         o_proj_slices = self.o_proj.weight.split(
@@ -182,103 +152,4 @@ def xformers_forward(
     else:
         attn_output = self.o_proj(attn_output)
 
-    return attn_output, attn_weights, past_key_value
-
-
-def sdp_attention_forward(
-    self,
-    hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
-    output_attentions: bool = False,
-    use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    # pylint: disable=duplicate-code
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    (
-        query_states,
-        key_states,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
-        query_states, key_states, cos, sin, position_ids
-    )
-    # [bsz, nh, t, hd]
-
-    if past_key_value is not None:
-        # reuse k, v, self_attention
-        key_states = torch.cat([past_key_value[0], key_states], dim=2)
-        value_states = torch.cat([past_key_value[1], value_states], dim=2)
-
-    past_key_value = (key_states, value_states) if use_cache else None
-
-    # We only apply sdp attention if we don't need to output the whole attention matrix
-    if not output_attentions:
-        with torch.backends.cuda.sdp_kernel():
-            attn_output = torch.nn.functional.scaled_dot_product_attention(
-                query_states,
-                key_states,
-                value_states,
-                attn_mask=attention_mask,
-                is_causal=False,
-            )
-            attn_weights = None
-    else:
-        attn_weights = torch.matmul(
-            query_states, key_states.transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
-
-        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
-            raise ValueError(
-                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
-                f" {attn_weights.size()}"
-            )
-
-        if attention_mask is not None:
-            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
-                raise ValueError(
-                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
-                )
-            attn_weights = attn_weights + attention_mask
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(
-            attn_weights, dim=-1, dtype=torch.float32
-        ).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
-
-    attn_output = attn_output.transpose(1, 2)
-    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-
-    attn_output = self.o_proj(attn_output)
-
-    return attn_output, attn_weights, past_key_value
+    return attn_output, None, past_key_value

src/axolotl/monkeypatch/relora.py

@@ -0,0 +1,393 @@
+"""Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
+import glob
+import json
+import logging
+import os.path
+import shutil
+from pathlib import Path
+from typing import Dict, List, Sequence
+
+import bitsandbytes as bnb
+import peft
+import safetensors.torch as st
+import torch
+from huggingface_hub import snapshot_download
+from torch.optim.lr_scheduler import LRScheduler
+from torch.optim.optimizer import Optimizer
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.relora")
+
+
+def reset_optimizer(optimizer: torch.optim.Optimizer):
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            param_state = optimizer.state[param]
+            for key in param_state:
+                if "qmap" in key:
+                    continue
+
+                if key == "step" and isinstance(param_state[key], int):
+                    param_state[key] = 0
+                else:
+                    param_state[key] = torch.zeros_like(param_state[key])
+
+
+class ReLoRACallback(TrainerCallback):
+    """Callback to merge LoRA weights into the base model and save full-weight checkpoints"""
+
+    def __init__(self, cfg: DictDefault):
+        self.relora_steps = cfg.relora_steps
+        self.cpu_offload = cfg.relora_cpu_offload
+        self.quantized = cfg.load_in_4bit or cfg.load_in_8bit
+        self.last_full_model = cfg.base_model
+        self.resume_from_checkpoint = cfg.resume_from_checkpoint
+
+        if not os.path.exists(self.last_full_model):
+            self.last_full_model = str(Path(snapshot_download(cfg.base_model)))
+
+        assert os.path.exists(
+            self.last_full_model
+        ), "for ReLORA base_model must be a local path"
+
+        self.num_lora_restarts = 0
+        self.need_full_save = False
+
+    def on_train_begin(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.resume_from_checkpoint:
+            weight_path = os.path.join(self.resume_from_checkpoint, "relora")
+            if not os.path.exists(weight_path):
+                LOG.warning(
+                    "Resuming ReLoRA from checkpoint, but no full-weight save found"
+                )
+            else:
+                LOG.info(f"Loading adjusted base weights from {weight_path}")
+                load_weight_checkpoint(model, weight_path)
+        return control
+
+    def on_step_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        optimizer: torch.optim.Optimizer,
+        **_kwargs,
+    ):
+        if state.global_step > 0 and state.global_step % self.relora_steps == 0:
+            checkpoint_folder = os.path.join(
+                args.output_dir,
+                f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}",
+                "relora",
+            )
+
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    checkpoint_folder,
+                    reinit=True,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+                reset_optimizer(optimizer)
+
+            if self.quantized:
+                self.last_full_model = checkpoint_folder
+            self.num_lora_restarts += 1
+
+        return control
+
+    def on_save(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        checkpoint_folder = os.path.join(
+            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}", "relora"
+        )
+        if (
+            state.global_step >= self.relora_steps
+            and state.global_step % self.relora_steps != 0
+        ):
+            if self.quantized:
+                if is_main_process() and self.last_full_model != checkpoint_folder:
+                    # ensure the latest full parameter save is in the latest checkpoint
+                    # folder, so that automatic pruning of checkpoints does not remove it
+                    LOG.info(f"moving last full parameter save to {checkpoint_folder}")
+                    os.makedirs(checkpoint_folder, exist_ok=True)
+                    chunks = glob.glob(
+                        f"{self.last_full_model}/model*.safetensors"
+                    ) + glob.glob(f"{self.last_full_model}/model*.index.json")
+                    for path in chunks:
+                        new_path = os.path.abspath(shutil.move(path, checkpoint_folder))
+                        try:
+                            os.symlink(new_path, path)
+                        except OSError:
+                            # probably on windows without permission to symlink
+                            pass
+
+                    self.last_full_model = checkpoint_folder
+            else:
+                model.model.save_pretrained(checkpoint_folder, safe_serialization=True)
+
+        return control
+
+    def on_log(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        logs: Dict[str, float],
+        **_kwargs,
+    ):
+        logs["num_lora_restarts"] = self.num_lora_restarts
+        return control
+
+    def on_train_end(
+        self,
+        args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.quantized:
+            # perform final merge and save
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    args.output_dir,
+                    reinit=False,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+        # no need to save if unquantized, as finetune.py will call merge_and_unload()
+        return control
+
+
+class ReLoRAScheduler(LRScheduler):
+    """Wraps another scheduler to apply per-lora-restart learning rate warmups."""
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        inner_schedule: LRScheduler,
+        relora_steps: int,
+        warmup_steps: int,
+        min_lr_scale: float = 0.001,
+    ) -> None:
+        self.inner_schedule = inner_schedule
+        self.relora_steps = relora_steps
+        self.warmup_steps = warmup_steps
+        self.min_lr_scale = min_lr_scale
+        super().__init__(optimizer, inner_schedule.last_epoch, inner_schedule.verbose)
+
+    def get_lr(self) -> float:
+        self.inner_schedule.last_epoch = self.last_epoch
+
+        original = self.inner_schedule.get_lr()
+        step = self.last_epoch
+        if step < self.relora_steps:
+            scale = 1
+        else:
+            cycle_t = min(1.0, (step % self.relora_steps) / self.warmup_steps)
+            scale = cycle_t * (1 - self.min_lr_scale) + self.min_lr_scale
+
+        if isinstance(original, Sequence):
+            return [lr * scale for lr in original]
+        return original * scale
+
+
+def sharded_paths(path: str, module_names: List[str]) -> Dict[str, str]:
+    model_name = "model.safetensors"
+    if not os.path.exists(str(Path(path) / model_name)) and not os.path.exists(
+        str(Path(path) / f"{model_name}.index.json")
+    ):
+        model_name = "pytorch_model.bin"
+
+    index_path = str(Path(path) / f"{model_name}.index.json")
+    if os.path.exists(index_path):
+        with open(index_path, "r", encoding="utf-8") as file:
+            data = json.load(file)
+        return data["weight_map"]
+    return {(module_name + ".weight"): model_name for module_name in module_names}
+
+
+def lora_delta_weight(layer: peft.tuners.lora.LoraLayer, device) -> torch.Tensor:
+    if isinstance(layer, (peft.tuners.lora.Linear8bitLt, peft.tuners.lora.Linear4bit)):
+        adapter = layer.active_adapter
+        return (
+            peft.utils.transpose(
+                layer.lora_B[adapter].weight.detach().to(device)
+                @ layer.lora_A[adapter].weight.detach().to(device),
+                getattr(layer, "fan_in_fan_out", False),
+            )
+            * layer.scaling[adapter]
+        )
+
+    return layer.get_delta_weight().to(device)
+
+
+def find_lora_modules(model: peft.LoraModel) -> Dict[str, peft.tuners.lora.LoraLayer]:
+    modules: Dict[str, peft.tuners.lora.LoraLayer] = {}
+
+    key_list = [key for key, _ in model.model.named_modules() if "lora" not in key]
+    for key in key_list:
+        try:
+            # pylint: disable=protected-access
+            _parent, target, _target_name = peft.utils._get_submodules(model.model, key)
+        except AttributeError:
+            continue
+
+        if isinstance(target, peft.tuners.lora.LoraLayer):
+            modules[key] = target
+
+    return modules
+
+
+def update_weights(
+    target: peft.tuners.lora.LoraLayer, new_weight: torch.Tensor, reinit: bool, device
+):
+    if reinit:
+        for adapter_name in target.lora_A:
+            target.reset_lora_parameters(adapter_name)
+        for adapter_name in target.lora_embedding_A:
+            target.reset_lora_parameters(adapter_name)
+
+    if isinstance(target, peft.tuners.lora.Linear4bit):
+        # This could be faster, but the quantization of Linear4bit weights occurs
+        # when the module is moved from cpu to gpu. Without meddling *too* deeply in
+        # PEFT's innards or maintaining a duplicate of that codepath, this is good
+        # enough for now.
+        target.weight.quant_state = None
+        target.weight.data = new_weight.cpu()
+        target.to(device)
+    elif isinstance(target, peft.tuners.lora.Linear8bitLt):
+        target.weight = bnb.nn.Int8Params(new_weight, requires_grad=False).to(device)
+    else:
+        target.weight.data = new_weight.to(device)
+
+
+def merge_and_save(
+    model: peft.LoraModel,
+    model_src: str,
+    model_dst: str,
+    reinit: bool = False,
+    quantized: bool = False,
+    cpu_offload: bool = False,
+    actually_save: bool = True,
+):
+    modules = find_lora_modules(model)
+
+    if not quantized:
+        for module_name, target in modules.items():
+            update = target.get_delta_weight(target.active_adapter).detach()
+            target.weight.data += update
+
+            if reinit:
+                for adapter_name in target.lora_A:
+                    target.reset_lora_parameters(adapter_name)
+                for adapter_name in target.lora_embedding_A:
+                    target.reset_lora_parameters(adapter_name)
+        return
+
+    os.makedirs(model_dst, exist_ok=True)
+    shard_paths = sharded_paths(model_src, modules.keys())
+    out_shard_paths = {}
+
+    unique_shards = list(set(shard_paths.values()))
+    for shard_path in unique_shards:
+        out_tensors = {}
+        if shard_path.endswith(".safetensors"):
+            in_tensors = st.load_file(str(Path(model_src) / shard_path))
+        else:
+            in_tensors = torch.load(Path(model_src) / shard_path)
+            if "state_dict" in in_tensors:
+                in_tensors = in_tensors["state_dict"]
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            orig_weight = in_tensors[key]
+            old_dev = target.weight.device
+            math_dev = "cpu" if cpu_offload else old_dev
+
+            delta_weight = lora_delta_weight(target, math_dev)
+            new_weight = orig_weight.to(math_dev) + delta_weight
+            del delta_weight
+
+            if actually_save:
+                out_tensors[key] = new_weight.half().cpu()
+
+            update_weights(target, new_weight, reinit=reinit, device=old_dev)
+
+        if actually_save:
+            out_shard_name = shard_path
+            if out_shard_name.startswith("pytorch_model"):
+                out_shard_name = (
+                    out_shard_name.replace("pytorch_model", "model").rstrip(".bin")
+                    + ".safetensors"
+                )
+
+            for module_name in in_tensors:
+                if module_name not in out_tensors:
+                    out_tensors[module_name] = in_tensors[module_name].half()
+                out_shard_paths[module_name] = out_shard_name
+
+            shard_fn = str(Path(model_dst) / out_shard_name)
+            LOG.info(f"saving tensors to {shard_fn}")
+            st.save_file(out_tensors, shard_fn, metadata={"format": "pt"})
+
+        del in_tensors
+        del out_tensors
+        torch.cuda.empty_cache()
+
+    if actually_save and len(unique_shards) > 1:
+        with open(
+            str(Path(model_dst, "model.safetensors.index.json")), "w", encoding="utf-8"
+        ) as file:
+            json.dump({"metadata": {}, "weight_map": out_shard_paths}, file)
+
+
+def load_weight_checkpoint(model: peft.LoraModel, checkpoint_path: str):
+    modules = find_lora_modules(model)
+    shard_paths = sharded_paths(checkpoint_path, modules.keys())
+    unique_shards = list(set(shard_paths.values()))
+
+    for shard_path in unique_shards:
+        tensors = st.load_file(os.path.join(checkpoint_path, shard_path))
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            new_weight = tensors[key]
+            update_weights(
+                target, new_weight, reinit=False, device=target.weight.device
+            )

src/axolotl/prompt_strategies/__init__.py

@@ -2,8 +2,10 @@
 
 import importlib
 
+from axolotl.prompt_strategies.user_defined import UserDefinedDatasetConfig
 
-def load(strategy, tokenizer, cfg):
+
+def load(strategy, tokenizer, cfg, ds_cfg):
     try:
         load_fn = "load"
         if strategy.split(".")[-1].startswith("load_"):
@@ -11,6 +13,9 @@ def load(strategy, tokenizer, cfg):
             strategy = ".".join(strategy.split(".")[:-1])
         mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies")
         func = getattr(mod, load_fn)
-        return func(tokenizer, cfg)
+        load_kwargs = {}
+        if strategy == "user_defined":
+            load_kwargs["ds_cfg"] = UserDefinedDatasetConfig(**ds_cfg)
+        return func(tokenizer, cfg, **load_kwargs)
     except Exception:  # pylint: disable=broad-exception-caught
         return None

src/axolotl/prompt_strategies/alpaca_w_system.py

@@ -57,6 +57,8 @@ class SystemDataPrompter(AlpacaPrompter):
     Alpaca Style Prompter that uses system prompts from the dataset
     """
 
+    system_format: str = "### System:\n{system}\n\n"
+
     def build_prompt_w_system(
         self,
         system: str,

src/axolotl/prompt_strategies/metharme.py

@@ -0,0 +1,76 @@
+"""Module containing the MetharmenPromptTokenizingStrategy and MetharmePrompter class"""
+
+import logging
+from typing import Tuple
+
+from axolotl.prompt_tokenizers import InstructionPromptTokenizingStrategy
+from axolotl.prompters import AlpacaPrompter
+
+LOG = logging.getLogger("axolotl")
+
+IGNORE_TOKEN_ID = -100
+
+# pylint: disable=duplicate-code
+
+
+class MetharmePromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
+    """
+    Tokenizing strategy for the Metharme models
+    """
+
+    def parse_instruction_fields(self, prompt) -> Tuple[str, str, str]:
+        return (prompt["prompt"], "", prompt["generation"])
+
+    def _tokenize(
+        self,
+        prompt: str,
+        add_eos_token: bool = True,
+        strip_bos_token: bool = False,
+        num_eos_tokens: int = 3,
+    ):
+        result = self.tokenizer(
+            prompt,
+            truncation=True,
+            max_length=self.sequence_len,
+            padding=False,
+            return_tensors=None,
+        )
+        if len(result["input_ids"]) == 0:
+            LOG.warning("Tokenizer result is empty. You may want to audit your dataset")
+        # If there's already an EOS token there, subtract from the number added
+        if result["input_ids"][-1] == self.tokenizer.eos_token_id:
+            num_eos_tokens -= 1
+
+        if num_eos_tokens > 0 and add_eos_token and len(result["input_ids"]) > 0:
+            for _ in range(num_eos_tokens):
+                if len(result["input_ids"]) < self.sequence_len:
+                    result["input_ids"].append(self.tokenizer.eos_token_id)
+                    result["attention_mask"].append(1)
+
+        if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
+            result["input_ids"] = result["input_ids"][1:]
+            result["attention_mask"] = result["attention_mask"][1:]
+
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+
+class MetharmePrompter(AlpacaPrompter):
+    """
+    Prompter for the Metharme models.
+    """
+
+    system_prompt = ""
+    system_no_input_prompt = ""
+    system_format = ""
+    turn_format = "{instruction}"
+    turn_no_input_format = "{instruction}"
+
+    def __init__(self, *args, **kwargs):  # pylint: disable=super-init-not-called
+        pass
+
+
+def load(tokenizer, cfg):
+    return MetharmePromptTokenizingStrategy(
+        MetharmePrompter(), tokenizer, cfg.train_on_inputs, cfg.sequence_len
+    )

src/axolotl/prompt_strategies/user_defined.py

@@ -0,0 +1,98 @@
+"""
+User Defined prompts with configuration from the YML config
+"""
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Optional, Tuple
+
+from axolotl.prompt_strategies.alpaca_w_system import (
+    InstructionWSystemPromptTokenizingStrategy,
+    SystemDataPrompter,
+)
+
+
+@dataclass
+class UserDefinedDatasetConfig:
+    """
+    dataclass configuration representing a userdefined dataset type
+    """
+
+    system_prompt: str = ""
+    field_system: str = "system"
+    field_instruction: str = "instruction"
+    field_input: str = "input"
+    field_output: str = "output"
+    format: str = "{instruction} {input} "
+    no_input_format: str = "{instruction} "
+    system_format: str = "{system}"
+
+    def __getitem__(self, item):
+        return getattr(self, item)
+
+
+class UserDefinedPromptTokenizationStrategy(InstructionWSystemPromptTokenizingStrategy):
+    """
+    Prompt Tokenization Strategy for user defined prompts
+    """
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[UserDefinedDatasetConfig] = None):
+    if not ds_cfg:
+        raise ValueError("Missing dataset prompt configuration")
+
+    system_prompt = ""
+    if ds_cfg.system_prompt:
+        system_prompt = ds_cfg.system_prompt
+
+    def parse_instruction_fields(
+        field_instruction,
+        field_input,
+        field_output,
+        field_system,
+        system_prompt,
+        prompt,
+    ) -> Tuple[str, str, str, str]:
+        return (
+            prompt[field_instruction],
+            prompt[field_input] if field_input in prompt else "",
+            prompt[field_output] if field_output in prompt else "",
+            prompt[field_system] if field_system in prompt else system_prompt,
+        )
+
+    turn_format = ds_cfg.format
+    turn_no_input_format = ds_cfg.no_input_format
+    system_format = ds_cfg.system_format
+
+    class UserDefinedPrompter(SystemDataPrompter):
+        """
+        Prompter for user defined prompts
+        """
+
+        def match_prompt_style(self):
+            self.turn_format = turn_format
+            self.turn_no_input_format = turn_no_input_format
+            self.system_format = system_format
+
+    prompter = UserDefinedPrompter()
+
+    strat = UserDefinedPromptTokenizationStrategy(
+        prompter,
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+    setattr(
+        strat,
+        "parse_instruction_fields",
+        partial(
+            parse_instruction_fields,
+            ds_cfg.field_instruction,
+            ds_cfg.field_input,
+            ds_cfg.field_output,
+            ds_cfg.field_system,
+            system_prompt,
+        ),
+    )
+    return strat

src/axolotl/prompt_tokenizers.py

@@ -13,7 +13,7 @@ from axolotl.prompters import IGNORE_TOKEN_ID
 LOG = logging.getLogger("axolotl")
 
 IGNORE_INDEX = -100
-LLAMA_DEFAULT_PAD_TOKEN = "[PAD]"  # nosec
+LLAMA_DEFAULT_PAD_TOKEN = "<pad>"  # nosec
 LLAMA_DEFAULT_EOS_TOKEN = "</s>"  # nosec
 LLAMA_DEFAULT_BOS_TOKEN = "<s>"  # nosec
 LLAMA_DEFAULT_UNK_TOKEN = "<unk>"  # nosec
@@ -85,7 +85,11 @@ class PromptTokenizingStrategy(abc.ABC):
             result["input_ids"].append(self.tokenizer.eos_token_id)
             result["attention_mask"].append(1)
 
-        if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
+        if (
+            len(result["input_ids"]) > 0
+            and result["input_ids"][0] == self.tokenizer.bos_token_id
+            and strip_bos_token
+        ):
             result["input_ids"] = result["input_ids"][1:]
             result["attention_mask"] = result["attention_mask"][1:]
 

src/axolotl/prompters.py

@@ -26,7 +26,7 @@ class AlpacaPrompter:
 
     system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
     system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
-    system_format: str
+    system_format: str = "{system}"
     turn_format: str
     turn_no_input_format: str
     prompt_style: Optional[PromptStyle] = None
@@ -63,13 +63,17 @@ class AlpacaPrompter:
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
         if input:
-            res = self.system_prompt + self.turn_format.format(
-                instruction=instruction, input=input
-            )
+            res = (
+                self.system_format.format(system=self.system_prompt)
+                if self.system_prompt
+                else ""
+            ) + self.turn_format.format(instruction=instruction, input=input)
         else:
-            res = self.system_no_input_prompt + self.turn_no_input_format.format(
-                instruction=instruction
-            )
+            res = (
+                self.system_format.format(system=self.system_no_input_prompt)
+                if self.system_prompt
+                else ""
+            ) + self.turn_no_input_format.format(instruction=instruction)
         if output:
             res = f"{res}{output}"
         yield res

src/axolotl/utils/callbacks.py

@@ -1,9 +1,19 @@
 """Callbacks for Trainer class"""
 
+from __future__ import annotations
+
 import logging
 import os
+from typing import TYPE_CHECKING, Dict, List
 
+import evaluate
+import numpy as np
+import pandas as pd
+import torch
+import torch.distributed as dist
+from datasets import load_dataset
 from optimum.bettertransformer import BetterTransformer
+from tqdm import tqdm
 from transformers import (
     TrainerCallback,
     TrainerControl,
@@ -13,8 +23,19 @@ from transformers import (
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
 
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.distributed import (
+    barrier,
+    gather_scalar_from_all_ranks,
+    get_world_size,
+    is_main_process,
+    zero_first,
+)
+
+if TYPE_CHECKING:
+    from axolotl.utils.trainer import AxolotlTrainingArguments
 
 LOG = logging.getLogger("axolotl.callbacks")
+IGNORE_INDEX = -100
 
 
 class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-methods
@@ -33,7 +54,9 @@ class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-
         )
 
         peft_model_path = os.path.join(checkpoint_folder, "adapter_model")
-        kwargs["model"].save_pretrained(peft_model_path)
+        kwargs["model"].save_pretrained(
+            peft_model_path, save_safetensors=args.save_safetensors
+        )
 
         return control
 
@@ -94,3 +117,192 @@ class GPUStatsCallback(
             log_gpu_memory_usage(LOG, "while training", self.cfg.device)
             self.logged = True
         return control
+
+
+def bench_eval_callback_factory(trainer, tokenizer):
+    accuracy = evaluate.load("accuracy")
+    abcd_idx = [
+        tokenizer("A", add_special_tokens=False).input_ids[0],
+        tokenizer("B", add_special_tokens=False).input_ids[0],
+        tokenizer("C", add_special_tokens=False).input_ids[0],
+        tokenizer("D", add_special_tokens=False).input_ids[0],
+        tokenizer("E", add_special_tokens=False).input_ids[0],
+        tokenizer("F", add_special_tokens=False).input_ids[0],
+        tokenizer("G", add_special_tokens=False).input_ids[0],
+    ]
+    bench_split = "eval"
+
+    def transform_bench_subject(example):
+        # Split on ':' and trim whitespace
+        parts = example["subject"].split(":")
+        first_part = (
+            parts[0].strip().lower().replace("-", "_")
+        )  # Lowercase the first part
+        second_part = (
+            parts[1].strip().replace("-", "_") if len(parts) > 1 else "all"
+        )  # Replace hyphens with underscores
+
+        # Return the transformed values
+        return {"name": first_part, "subject": second_part}
+
+    if trainer.args.bench_dataset == "mmlu-zs":
+        bench_dataset = load_dataset(
+            "openaccess-ai-collective/mmlu-evals",
+            data_files={
+                "eval": "zero_shot_mmlu_val.json",
+                "test": "zero_shot_mmlu_test.json",
+            },
+        )
+        # bench_dataset = bench_dataset.remove_columns("subject")
+    # MMLU Five-shot (Eval/Test only)
+    elif trainer.args.bench_dataset in ["mmlu", "mmlu-fs"]:
+        bench_dataset = load_dataset(
+            "openaccess-ai-collective/mmlu-evals",
+            data_files={
+                "eval": "five_shot_mmlu_val.json",
+                "test": "five_shot_mmlu_test.json",
+            },
+        )
+        # bench_dataset = bench_dataset.remove_columns('subject')
+    elif "/" in trainer.args.bench_dataset:
+        bench_ds = trainer.args.bench_dataset
+        bench_ds_name = "/".join(bench_ds.split("/", 2)[:2])
+        bench_ds_data_file = "/".join(bench_ds.split("/", 2)[2:])
+        bench_dataset = load_dataset(
+            bench_ds_name,
+            data_files={
+                "eval": bench_ds_data_file,
+            },
+        )
+        bench_dataset["eval"] = bench_dataset["eval"].map(transform_bench_subject)
+    else:
+        raise ValueError(
+            f"unhandled value `{trainer.args.bench_dataset}` for bench_dataset training args"
+        )
+    bench_dataset = bench_dataset[trainer.args.bench_split]
+    if trainer.args.max_bench_samples is not None:
+        bench_dataset = bench_dataset.select(range(trainer.args.max_bench_samples))
+
+    def tokenize_evals(example):
+        source = f"{tokenizer.bos_token}{example['input']}"
+        target = f"{example['output']}{tokenizer.eos_token}"
+
+        tokenized_source = tokenizer(
+            source,
+            max_length=2048,
+            truncation=True,
+            add_special_tokens=False,
+        )
+        tokenized_target = tokenizer(
+            target,
+            max_length=2048,
+            truncation=True,
+            add_special_tokens=False,
+        )
+        input_ids = tokenized_source["input_ids"] + tokenized_target["input_ids"]
+        labels = [IGNORE_INDEX] * len(tokenized_source["input_ids"]) + tokenized_target[
+            "input_ids"
+        ]
+
+        return {
+            "input_ids": input_ids,
+            "labels": labels,
+            "subject": example["subject"],
+        }
+
+    with zero_first(is_main_process()):
+        bench_dataset = bench_dataset.map(tokenize_evals)
+        bench_dataset = bench_dataset.filter(lambda x: x["labels"][-2] in abcd_idx)
+
+    class BenchEvalCallback(TrainerCallback):
+        """
+        TrainerCallback that runs the MMLU evals
+        """
+
+        def on_evaluate(
+            self,
+            args: AxolotlTrainingArguments,
+            state: TrainerState,  # pylint: disable=unused-argument
+            control: TrainerControl,  # pylint: disable=unused-argument
+            metrics: Dict[str, float],  # pylint: disable=unused-argument
+            **kwargs,  # pylint: disable=unused-argument
+        ):
+            data_loader = trainer.get_bench_dataloader(
+                bench_dataset.remove_columns(["input", "subject", "output", "name"])
+            )
+            trainer.model.eval()
+            preds, refs = [], []
+            loss_bench = 0
+            for batch in tqdm(data_loader, total=len(data_loader)):
+                (loss, logits, labels) = trainer.prediction_step(
+                    trainer.model,
+                    batch,
+                    prediction_loss_only=False,
+                )
+                # There are two tokens, the output, and eos token.
+                for i, logit in enumerate(logits):
+                    label_non_zero_id = (batch["labels"][i] != IGNORE_INDEX).nonzero()[
+                        0
+                    ][0]
+                    logit_abcd = logit[label_non_zero_id - 1][abcd_idx]
+                    preds.append(torch.argmax(logit_abcd).item())
+                labels = labels[labels != IGNORE_INDEX].view(-1, 2)[:, 0]
+                refs += [
+                    abcd_idx.index(label) if label in abcd_idx else -1
+                    for label in labels.tolist()
+                ]
+                loss_bench += loss.item()
+            # Extract results by subject.
+            bench_name = bench_dataset["name"]
+            bench_names: dict = {s: {"refs": [], "preds": []} for s in set(bench_name)}
+            for s, p, r in zip(bench_name, preds, refs):  # pylint: disable=invalid-name
+                bench_names[s]["preds"].append(p)
+                bench_names[s]["refs"].append(r)
+            barrier()
+            local_bench_names = bench_names
+            gathered_bench_names: List[Dict] = [{} for _ in range(get_world_size())]
+            # Gather results from all GPUs to GPU 0
+
+            loss_bench_ranks = gather_scalar_from_all_ranks(
+                lambda: loss_bench, get_world_size()
+            )
+            len_data_loader_ranks = gather_scalar_from_all_ranks(
+                lambda: len(data_loader), get_world_size()
+            )
+
+            if not is_main_process():
+                dist.gather_object(local_bench_names, dst=0)
+            else:
+                dist.gather_object(local_bench_names, gathered_bench_names, dst=0)
+                bench_loss = sum(loss_bench_ranks) / sum(len_data_loader_ranks)
+                results = {"bench_loss": bench_loss}
+
+                # Combine results from all GPUs
+                combined_bench_names: Dict[str, Dict[str, List]] = {}
+                for bench_name in gathered_bench_names:
+                    for name, data in bench_name.items():
+                        if name not in combined_bench_names:
+                            combined_bench_names[name] = {"refs": [], "preds": []}
+                        combined_bench_names[name]["refs"].extend(data["refs"])
+                        combined_bench_names[name]["preds"].extend(data["preds"])
+
+                bench_scores = []
+                for (
+                    bench_name
+                ) in combined_bench_names:  # pylint: disable=consider-using-dict-items
+                    bench_score = accuracy.compute(
+                        references=combined_bench_names[bench_name]["refs"],
+                        predictions=combined_bench_names[bench_name]["preds"],
+                    )["accuracy"]
+                    if not pd.isna(bench_score):
+                        results[
+                            f"bench_{bench_split}_accuracy_{bench_name}"
+                        ] = bench_score
+                        bench_scores.append(bench_score)
+                    else:
+                        results[f"bench_{bench_split}_accuracy_{bench_name}"] = 0.0
+                        bench_scores.append(0.0)
+                results[f"bench_{bench_split}_accuracy"] = np.mean(bench_scores)
+                trainer.log(results)
+
+    return BenchEvalCallback

src/axolotl/utils/config.py

@@ -62,6 +62,13 @@ def normalize_config(cfg):
     else:
         torch.backends.cuda.matmul.allow_tf32 = cfg.tf32 or False
 
+    if cfg.bf16 or cfg.bfloat16:
+        cfg.torch_dtype = torch.bfloat16
+    elif cfg.load_in_8bit or cfg.fp16 or cfg.float16:
+        cfg.torch_dtype = torch.float16
+    else:
+        cfg.torch_dtype = torch.float32
+
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
 
@@ -119,6 +126,19 @@ def validate_config(cfg):
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
+    if cfg.relora_steps:
+        if cfg.adapter not in ("lora", "qlora"):
+            raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
+
+        if cfg.fsdp:
+            raise ValueError("fsdp not supported with ReLoRA")
+
+        if cfg.deepspeed:
+            raise ValueError("deepspeed not supported with ReLoRA")
+
+        if cfg.lr_scheduler == "one_cycle":
+            raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
+
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."

src/axolotl/utils/data.py

@@ -41,6 +41,7 @@ from axolotl.prompters import (
     ShareGPTPrompter,
     SummarizeTLDRPrompter,
 )
+from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.trainer import (
     calculate_total_num_steps,
@@ -53,9 +54,10 @@ DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
 
 def prepare_dataset(cfg, tokenizer):
     if not cfg.pretraining_dataset:
-        train_dataset, eval_dataset = load_prepare_datasets(
-            tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
-        )
+        with zero_first(is_main_process()):
+            train_dataset, eval_dataset = load_prepare_datasets(
+                tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
+            )
     else:
         train_dataset = load_pretraining_dataset(
             cfg.pretraining_dataset,
@@ -160,8 +162,15 @@ def load_tokenized_prepared_datasets(
                         split=None,
                     )
                 elif local_path.is_file():
+                    ds_type = "json"
+                    if d.ds_type:
+                        ds_type = d.ds_type
+                    elif ".parquet" in d.path:
+                        ds_type = "parquet"
+                    elif ".arrow" in d.path:
+                        ds_type = "arrow"
                     ds = load_dataset(
-                        "json",
+                        ds_type,
                         name=d.name,
                         data_files=d.path,
                         streaming=False,
@@ -198,13 +207,27 @@ def load_tokenized_prepared_datasets(
                     )
                 else:
                     ds = ds.shuffle(seed=seed).shard(num_shards=d.shards, index=0)
+
+            d_base_type = d_prompt_style = None
             d_type = d.type
-            d_type_split = d_type.split(":")
-            d_base_type = d_type_split[0]
-            d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
+            if isinstance(d_type, str):
+                d_type_split = d_type.split(":")
+                d_base_type = d_type_split[0]
+                d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
             if "train" in ds:
                 ds = ds["train"]
-            if ds_strategy := load(d.type, tokenizer, cfg):
+            if (
+                "input_ids" in ds.features
+                and "attention_mask" in ds.features
+                and "labels" in ds.features
+            ):
+                # dataset is already tokenized, just drop it straight in
+                datasets.append(ds)
+            elif isinstance(d.type, DictDefault):
+                ds_strategy = load("user_defined", tokenizer, cfg, d.type.to_dict())
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif ds_strategy := load(d.type, tokenizer, cfg, d):
                 ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
                 datasets.append(ds_wrapper)
             elif d_base_type == "alpaca":

src/axolotl/utils/dataloader.py

@@ -243,6 +243,18 @@ class MultipackDistributedDataloader:
             len_remaining -= 1
             if not len_remaining:
                 return
+        # yield a no-op for cases where we don't have any data left to pack
+        for i in range(0, len_remaining):
+            yield self.collate_fn(
+                [
+                    {
+                        "input_ids": [0],
+                        "labels": [-100],
+                        "attention_mask": [True],
+                        "position_ids": [0],
+                    }
+                ]
+            )
 
     def _len_est(self):
         lengths_sum = np.sum(self.lengths)

src/axolotl/utils/distributed.py

@@ -1,8 +1,10 @@
 """
 utility helpers for distributed checks
 """
+import os
 from contextlib import contextmanager
 
+import torch
 import torch.distributed as dist
 from accelerate import Accelerator
 
@@ -43,6 +45,10 @@ def is_main_process():
     return dist.get_rank() == 0
 
 
+def get_world_size():
+    return int(os.getenv("WORLD_SIZE", "1"))
+
+
 @contextmanager
 def zero_first(is_main):
     """
@@ -53,3 +59,35 @@ def zero_first(is_main):
     yield
     if is_main:  # then rank 0 waits after it has run the context
         barrier()
+
+
+def gather_scalar_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-name
+    """
+    Run a callable 'fn' on all ranks and gather the results on the specified rank.
+
+    Args:
+    - fn (callable): A function that computes the value. This should not have any side effects.
+    - rank (int, optional): The rank that gathers the values. Default is 0.
+    - world_size (int, optional): Total number of processes in the current distributed setup.
+
+    Returns:
+    - A list of computed values from all ranks if on the gathering rank, otherwise None.
+    """
+    value_scalar = fn()
+    value_tensor = torch.tensor(value_scalar, device=dist.get_rank()).float()
+
+    if not is_main_process():
+        dist.gather(value_tensor, dst=0)
+    else:
+        gathered_tensors = [torch.zeros_like(value_tensor) for _ in range(world_size)]
+        dist.gather(value_tensor, gather_list=gathered_tensors, dst=0)
+
+        # Convert tensors back to their original type (int or float)
+        gathered_values = []
+        for tensor in gathered_tensors:
+            if tensor == tensor.int():
+                gathered_values.append(int(tensor.item()))
+            else:
+                gathered_values.append(float(tensor.item()))
+        return gathered_values
+    return None

src/axolotl/utils/models.py

@@ -21,7 +21,7 @@ from transformers import (  # noqa: F401
     PreTrainedTokenizerBase,
 )
 
-from axolotl.prompt_tokenizers import LLAMA_DEFAULT_PAD_TOKEN
+from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
 from axolotl.utils.bench import log_gpu_memory_usage
 
 LOG = logging.getLogger("axolotl")
@@ -54,11 +54,18 @@ def load_tokenizer(cfg):
         **tokenizer_kwargs,
     )
 
-    if tokenizer.__class__.__name__ in [
-        "LlamaTokenizer",
-        "LlamaTokenizerFast",
-    ]:
-        tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
+    if (
+        tokenizer.__class__.__name__
+        in [
+            "LlamaTokenizer",
+            "LlamaTokenizerFast",
+            "CodeLlamaTokenizer",
+        ]
+        and hasattr(tokenizer, "pad_token")
+        and not tokenizer.pad_token
+    ):
+        # set a pad_token, but use eos_token so we don't add a new token
+        tokenizer.pad_token = LLAMA_DEFAULT_EOS_TOKEN
 
     LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
     LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
@@ -103,7 +110,7 @@ def load_model(
             )
 
             LOG.info("patching with flash attention")
-            replace_llama_attn_with_flash_attn()
+            replace_llama_attn_with_flash_attn(packed=cfg.sample_packing)
     elif cfg.is_llama_derived_model and cfg.xformers_attention:
         from axolotl.monkeypatch.llama_attn_hijack_xformers import (
             hijack_llama_attention,
@@ -112,9 +119,7 @@ def load_model(
         LOG.info("patching with xformers attention")
         hijack_llama_attention()
     elif cfg.is_llama_derived_model and cfg.sdp_attention:
-        from axolotl.monkeypatch.llama_attn_hijack_xformers import (
-            hijack_llama_sdp_attention,
-        )
+        from axolotl.monkeypatch.llama_attn_hijack_sdp import hijack_llama_sdp_attention
 
         LOG.info("patching with sdp attention")
         hijack_llama_sdp_attention()
@@ -148,12 +153,6 @@ def load_model(
         LOG.info("patching _expand_mask")
         hijack_expand_mask()
 
-    if cfg.bf16 or cfg.bfloat16:
-        torch_dtype = torch.bfloat16
-    elif cfg.load_in_8bit or cfg.fp16 or cfg.float16:
-        torch_dtype = torch.float16
-    else:
-        torch_dtype = torch.float32
     try:
         if cfg.gptq:
             from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
@@ -185,7 +184,7 @@ def load_model(
             load_in_4bit=True,
             llm_int8_threshold=6.0,
             llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=torch_dtype,
+            bnb_4bit_compute_dtype=cfg.torch_dtype,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
@@ -244,7 +243,7 @@ def load_model(
                 device_map=cfg.device_map,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
+                torch_dtype=cfg.torch_dtype,
                 **model_kwargs,
             )
         # elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
@@ -279,7 +278,7 @@ def load_model(
                 device_map=cfg.device_map,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
+                torch_dtype=cfg.torch_dtype,
                 trust_remote_code=cfg.trust_remote_code or False,
                 **model_kwargs,
             )
@@ -310,7 +309,7 @@ def load_model(
                 device_map=cfg.device_map,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
+                torch_dtype=cfg.torch_dtype,
                 trust_remote_code=cfg.trust_remote_code or False,
                 **model_kwargs,
             )
@@ -324,7 +323,7 @@ def load_model(
             device_map=cfg.device_map,
             load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
             load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-            torch_dtype=torch_dtype,
+            torch_dtype=cfg.torch_dtype,
             trust_remote_code=cfg.trust_remote_code or False,
             **model_kwargs,
         )
@@ -349,6 +348,15 @@ def load_model(
     if model.device.type == "cuda":
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
+    # make sure these are fp32 per Ramesh et al. (2021)
+    for name, module in model.named_modules():
+        if "norm" in name:
+            module.to(torch.float32)
+        if "lm_head" in name or "embed_tokens" in name:
+            if hasattr(module, "weight"):
+                module.to(torch.float32)
+
+    needs_fa2_dtype = cfg.adapter or cfg.fsdp
     if not cfg.gptq and (
         (cfg.adapter == "lora" and load_in_8bit)
         or (cfg.adapter == "qlora" and cfg.load_in_4bit)
@@ -357,16 +365,18 @@ def load_model(
         model = prepare_model_for_kbit_training(
             model, use_gradient_checkpointing=cfg.gradient_checkpointing
         )
+        needs_fa2_dtype = True
 
-        # LlamaRMSNorm layers are in fp32 after kbit_training, so we need to
-        # convert them back to fp16/bf16 for flash-attn compatibility.
-        if cfg.flash_attention and cfg.is_llama_derived_model:
-            for name, module in model.named_modules():
-                if "norm" in name:
-                    module.to(torch_dtype)
-                if "lm_head" in name or "embed_tokens" in name:
-                    if hasattr(module, "weight"):
-                        module.to(torch_dtype)
+    # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
+    # convert them back to fp16/bf16 for flash-attn compatibility.
+    if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
+        LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
+        for name, module in model.named_modules():
+            if "norm" in name:
+                module.to(cfg.torch_dtype)
+            if "lm_head" in name or "embed_tokens" in name:
+                if hasattr(module, "weight"):
+                    module.to(cfg.torch_dtype)
 
     model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
@@ -440,7 +450,7 @@ def load_llama_adapter(model, cfg):
     )
 
     if cfg.lora_model_dir:
-        LOG.info("Loading pretained LORA")
+        LOG.debug("Loading pretained PEFT - llama_adapter")
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
@@ -502,6 +512,7 @@ def load_lora(model, cfg):
     )
 
     if cfg.lora_model_dir:
+        LOG.debug("Loading pretained PEFT - LoRA")
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,

src/axolotl/utils/trainer.py

@@ -10,28 +10,30 @@ from functools import partial
 from pathlib import Path
 from typing import Optional, Union
 
-import bitsandbytes as bnb
 import numpy as np
 import torch.cuda
 import transformers
 from datasets import Dataset, set_caching_enabled
-from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import DataLoader, DistributedSampler, RandomSampler
+from torch.utils.data import (
+    DataLoader,
+    DistributedSampler,
+    RandomSampler,
+    SequentialSampler,
+)
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
-from transformers.trainer_pt_utils import get_parameter_names
+from transformers.trainer_pt_utils import SequentialDistributedSampler
 
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     GPUStatsCallback,
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
+    bench_eval_callback_factory,
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
-from axolotl.utils.schedulers import (
-    InterpolatingLogScheduler,
-    get_cosine_schedule_with_quadratic_warmup,
-)
+from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
 LOG = logging.getLogger("axolotl")
 
@@ -124,6 +126,35 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=1,
         metadata={"help": "the multiplier for the max len for packed sequences"},
     )
+    relora_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to reset for ReLoRA"},
+    )
+    relora_warmup_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
+    )
+    bench_split: Optional[str] = field(
+        default="eval", metadata={"help": "The benchmark split to run on"}
+    )
+    bench_dataset: Optional[str] = field(
+        default="pharaouk/dharma-1/dharma_1_mini.json",
+        metadata={
+            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
+        },
+    )
+    do_bench_eval: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
+    )
+    max_bench_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
+        },
+    )
+    bench_source_max_len: int = field(
+        default=2048, metadata={"help": "Maximum source sequence length for bench."}
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -133,6 +164,10 @@ class AxolotlTrainer(Trainer):
 
     args = None  # type: AxolotlTrainingArguments
 
+    def __init__(self, *args, bench_data_collator=None, **kwargs):
+        self.bench_data_collator = bench_data_collator
+        super().__init__(*args, **kwargs)
+
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
     ):
@@ -171,6 +206,18 @@ class AxolotlTrainer(Trainer):
             )
         return super()._get_train_sampler()
 
+    def _get_eval_sampler(
+        self, eval_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size > 1 and self.args.sample_packing:
+            return SequentialDistributedSampler(
+                eval_dataset,
+                num_replicas=self.args.world_size,
+                rank=self.args.process_index,
+                batch_size=self.args.per_device_eval_batch_size,
+            )
+        return super()._get_eval_sampler(eval_dataset)
+
     def get_train_dataloader(self) -> Union[DataLoader, MultipackDistributedDataloader]:
         if self.args.sample_packing:
             train_sampler = self._get_train_sampler()
@@ -195,6 +242,7 @@ class AxolotlTrainer(Trainer):
             eval_dataset = (
                 eval_dataset if eval_dataset is not None else self.eval_dataset
             )
+
             eval_sampler = self._get_eval_sampler(eval_dataset)
             return self.accelerator.prepare(
                 MultipackDistributedDataloader(
@@ -210,6 +258,31 @@ class AxolotlTrainer(Trainer):
             )
         return super().get_eval_dataloader(eval_dataset)
 
+    def _get_bench_sampler(
+        self, bench_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size <= 1:
+            return SequentialSampler(bench_dataset)
+        return None
+
+    def get_bench_dataloader(
+        self,
+        bench_dataset: Dataset,
+    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+        dataloader_params = {
+            "batch_size": self.args.eval_batch_size,
+            "collate_fn": self.bench_data_collator,
+            "num_workers": self.args.dataloader_num_workers,
+            "pin_memory": self.args.dataloader_pin_memory,
+        }
+
+        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
+            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
+            dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+        return DataLoader(bench_dataset, **dataloader_params)
+        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
+
     def compute_loss(self, model, inputs, return_outputs=False):
         # use one's weighted cross entropy loss calc
         # if self.args.sample_packing:
@@ -249,6 +322,39 @@ class OneCycleLRSchedulerTrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
+class ReLoRATrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
 def add_position_ids(sample):
     sample["position_ids"] = torch.arange(len(sample["input_ids"]))
     return sample
@@ -268,15 +374,15 @@ def disable_datasets_caching():
 
 
 def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
+    drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
+    train_dataset = train_dataset.filter(drop_long, num_proc=os.cpu_count())
+    if eval_dataset:
+        eval_dataset = eval_dataset.filter(drop_long, num_proc=os.cpu_count())
+
     if cfg.sample_packing:
-        drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
-        train_dataset = train_dataset.filter(drop_long, num_proc=os.cpu_count()).map(
-            add_position_ids, num_proc=os.cpu_count()
-        )
+        train_dataset = train_dataset.map(add_position_ids, num_proc=os.cpu_count())
         if eval_dataset:
-            eval_dataset = eval_dataset.filter(drop_long, num_proc=os.cpu_count()).map(
-                add_position_ids, num_proc=os.cpu_count()
-            )
+            eval_dataset = eval_dataset.map(add_position_ids, num_proc=os.cpu_count())
     return train_dataset, eval_dataset
 
 
@@ -355,10 +461,16 @@ def calculate_total_num_steps(cfg, train_dataset, tokenizer):
 
 def setup_fsdp_envs(cfg):
     os.environ["ACCELERATE_USE_FSDP"] = "true"
+    if cfg.fsdp_config.fsdp_offload_params:
+        os.environ["FSDP_OFFLOAD_PARAMS"] = "true"
     if cfg.fsdp_config.fsdp_sync_module_states:
         os.environ["FSDP_SYNC_MODULE_STATES"] = "true"
     if cfg.fsdp_config.fsdp_state_dict_type:
         os.environ["FSDP_STATE_DICT_TYPE"] = cfg.fsdp_config.fsdp_state_dict_type
+    if cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap:
+        os.environ[
+            "FSDP_TRANSFORMER_CLS_TO_WRAP"
+        ] = cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap
 
 
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
@@ -455,6 +567,18 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         # we have an eval set, but no steps defined, use epoch
         training_arguments_kwargs["evaluation_strategy"] = "epoch"
 
+    if cfg.save_strategy:
+        training_arguments_kwargs["save_strategy"] = cfg.save_strategy
+    else:
+        training_arguments_kwargs["save_strategy"] = (
+            "steps" if cfg.save_steps else "epoch"
+        )
+
+    if cfg.do_bench_eval:
+        training_arguments_kwargs["do_bench_eval"] = cfg.do_bench_eval
+        if cfg.bench_dataset:
+            training_arguments_kwargs["bench_dataset"] = cfg.bench_dataset
+
     training_args = AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
         max_steps=total_num_steps if cfg.max_steps else -1,
         max_seq_length=cfg.sequence_len,
@@ -466,7 +590,6 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         eval_accumulation_steps=cfg.gradient_accumulation_steps,
         num_train_epochs=cfg.num_epochs,
         learning_rate=cfg.learning_rate,
-        save_strategy="steps" if cfg.save_steps else "epoch",
         save_steps=cfg.save_steps,
         output_dir=cfg.output_dir,
         save_total_limit=cfg.save_total_limit if cfg.save_total_limit else 4,
@@ -489,6 +612,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
         sample_packing=cfg.sample_packing if cfg.sample_packing else False,
         sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+        relora_steps=cfg.relora_steps,
+        relora_warmup_steps=cfg.relora_warmup_steps,
         **training_arguments_kwargs,
     )
 
@@ -498,69 +623,13 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         if Path(cfg.torchdistx_path).exists():
             sys.path.append(cfg.torchdistx_path)
             importlib.import_module("torchdistx")
-    if (
-        cfg.optimizer == "adamw_bnb_8bit"
-        and not cfg.gptq
-        and "deepspeed" not in training_arguments_kwargs
-        and not cfg.fsdp
-    ):
-        decay_parameters = get_parameter_names(model, [nn.LayerNorm])
-        decay_parameters = [name for name in decay_parameters if "bias" not in name]
-        optimizer_grouped_parameters = [
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": training_args.weight_decay,
-            },
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n not in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": 0.0,
-            },
-        ]
-
-        optimizer = bnb.optim.Adam8bit(
-            optimizer_grouped_parameters,
-            betas=(training_args.adam_beta1, training_args.adam_beta2),
-            eps=training_args.adam_epsilon,
-            lr=training_args.learning_rate,
-        )
-
-        if cfg.lr_scheduler == "one_cycle":
-            lr_scheduler_kwargs = (
-                cfg.lr_scheduler_kwargs if cfg.lr_scheduler_kwargs else {}
-            )
-            lr_scheduler = OneCycleLR(
-                optimizer,
-                cfg.learning_rate,
-                total_steps=total_num_steps,
-                epochs=cfg.num_epochs,
-                div_factor=cfg.lr_div_factor if cfg.lr_div_factor else 6,
-                **lr_scheduler_kwargs,
-            )
-        elif cfg.lr_scheduler == "log_sweep":
-            lr_scheduler = InterpolatingLogScheduler(
-                optimizer,
-                cfg.warmup_steps,
-                cfg.log_sweep_min_lr if cfg.log_sweep_min_lr else 1e-10,
-                cfg.log_sweep_max_lr if cfg.log_sweep_max_lr else 10,
-            )
-        else:
-            lr_scheduler = transformers.get_cosine_schedule_with_warmup(
-                optimizer,
-                training_args.warmup_steps,
-                total_num_steps,
-            )
-        trainer_kwargs["optimizers"] = (optimizer, lr_scheduler)
 
     callbacks = []
     callbacks.append(GPUStatsCallback(cfg))
+
+    if cfg.relora_steps:
+        callbacks.append(ReLoRACallback(cfg))
+
     # TODO on_save callback to sync checkpoints to GCP/AWS in background
     if cfg.early_stopping_patience:
         early_stop_cb = EarlyStoppingCallback(
@@ -605,11 +674,11 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
                 num_proc=32,
             )
 
-    trainer_cls = (
-        OneCycleLRSchedulerTrainer
-        if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora")
-        else AxolotlTrainer
-    )
+    trainer_cls = AxolotlTrainer
+    if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora"):
+        trainer_cls = OneCycleLRSchedulerTrainer
+    elif cfg.relora_steps:
+        trainer_cls = ReLoRATrainer
     trainer = trainer_cls(
         model=model,
         train_dataset=train_dataset,
@@ -620,8 +689,16 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
             return_tensors="pt",
             **data_collator_kwargs,
         ),
+        bench_data_collator=transformers.DataCollatorForSeq2Seq(
+            tokenizer,
+            return_tensors="pt",
+            **data_collator_kwargs,
+        ),
         callbacks=callbacks,
         **trainer_kwargs,
     )
 
+    if cfg.do_bench_eval:
+        trainer.add_callback(bench_eval_callback_factory(trainer, tokenizer))
+
     return trainer