add kernels for gpt oss models (#3020)

* add kernels for gpt oss models

* add support for gpt-oss

* typo incorrect package

* fix: layout for configs and added wandb/epochs

* add gptoss example w offload and set moe leaf for z3

* add support for Mxfp4Config from yaml

* update yaml to use official model

* fix lora and don't allow triton to go above 3.3.1

* fix lr and tweak vram use

* fix range for triton since pinned wasn't compatible with toch 2.6.0

* update cce with gpt oss patches

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -40,7 +40,7 @@
         "%%capture\n",
         "# This step can take ~5-10 minutes to install dependencies\n",
         "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-        "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@cbd58e0\""
+        "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@48b5169\""
       ]
     },
     {

examples/gpt-oss/README.md

@@ -0,0 +1,9 @@
+# OpenAI's GPT-OSS
+
+GPT-OSS is a 20 billion parameter MoE model trained by OpenAI, released in August 2025.
+
+- 20B Full Parameter SFT can be trained on 8x48GB GPUs (peak reserved memory @ ~36GiB/GPU) - [YAML](./gpt-oss-20b-fft-fsdp2.yaml)
+- 20B LoRA SFT (all linear layers, and experts in last two layers) can be trained a single GPU (peak reserved memory @ ~47GiB)
+  - removing the experts from `lora_target_parameters` will allow the model to fit around ~44GiB of VRAM
+  - [YAML](./gpt-oss-20b-sft-lora-singlegpu.yaml)
+- 20B Full Parameter SFT with FSDP2 offloading can be trained on 2x24GB GPUs (peak reserved memory @ ~21GiB/GPU) - [YAML](./gpt-oss-20b-fft-fsdp2-offload.yaml)

examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml

@@ -0,0 +1,62 @@
+base_model: openai/gpt-oss-20b
+use_kernels: true
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: winglian/pirate-ultrachat-10k
+    type: chat_template
+    split: train
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: true
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: GptOssDecoderLayer
+  reshard_after_forward: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml

@@ -0,0 +1,62 @@
+base_model: openai/gpt-oss-20b
+use_kernels: true
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: winglian/pirate-ultrachat-10k
+    type: chat_template
+    split: train
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_8bit
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: GptOssDecoderLayer
+  reshard_after_forward: true

examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml

@@ -0,0 +1,64 @@
+base_model: openai/gpt-oss-20b
+use_kernels: true
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by not putting model to GPU before sharding
+
+datasets:
+  - path: winglian/pirate-ultrachat-10k
+    type: chat_template
+    split: train
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+adapter: lora
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.0  # dropout not supported when using LoRA over expert parameters
+lora_target_linear: true
+lora_target_parameters:  # target the experts in the last two layers
+  - "22._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
+  - "22._checkpoint_wrapped_module.mlp.experts.down_proj"
+  - "23._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
+  - "23._checkpoint_wrapped_module.mlp.experts.down_proj"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_8bit
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-4
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+warmup_ratio: 0.1
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"