add 120b and deepspeed zero3 examples (#3035) [skip ci]

* add 120b and deepspeed zero3 examples

* add a bit of flavor and cleanup gpt oss readme

* fix: remove expert vram usage

* fix: remove redundant EOS token from eot_tokens

* feat: add 120B to docs

---------

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

examples/gpt-oss/README.md

@@ -16,11 +16,10 @@ pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Choose one of the following configs below for training the 20B model.
+2. Choose one of the following configs below for training the 20B model. (for 120B, see [below](#training-120b))
 
 ```bash
-# LoRA SFT linear layers & 2 experts (1x48GB @ ~47GiB)
-# (only linear layers @ ~44GiB)
+# LoRA SFT linear layers (1x48GB @ ~44GiB)
 axolotl train examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
 
 # FFT SFT with offloading (2x24GB @ ~21GiB/GPU)
@@ -30,9 +29,16 @@ axolotl train examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
 axolotl train examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
 ```
 
-Notes:
-- 120B coming soon!
-- Memory usage taken from `device_mem_reserved(gib)` from logs.
+Note: Memory usage taken from `device_mem_reserved(gib)` from logs.
+
+### Training 120B
+
+On 8xH100s
+
+```bash
+# FFT SFT with offloading (8x80GB @ ~49GiB/GPU)
+axolotl train examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
+```
 
 ### Tool use
 

examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -0,0 +1,67 @@
+# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
+# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
+base_model: axolotl-ai-co/gpt-oss-120b-dequantized
+
+use_kernels: false
+
+dp_shard_size: 16  # requires 2x8xH100 nodes
+
+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: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 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: 0.03
+
+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
+  cpu_ram_efficient_loading: true

examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml

@@ -0,0 +1,58 @@
+base_model: openai/gpt-oss-20b
+use_kernels: false
+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: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+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: 0.03
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+# choose the zero3 configuration that best fits your system capabilities
+deepspeed: deepspeed_configs/zero3_bf16.json

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

@@ -54,7 +54,6 @@ warmup_ratio: 0.03
 special_tokens:
 eot_tokens:
   - "<|end|>"
-  - "<|return|>"
 
 fsdp_version: 2
 fsdp_config:
@@ -63,4 +62,7 @@ fsdp_config:
   auto_wrap_policy: TRANSFORMER_BASED_WRAP
   transformer_layer_cls_to_wrap: GptOssDecoderLayer
   reshard_after_forward: true
-#  cpu_ram_efficient_loading: true
+  #  cpu_ram_efficient_loading: true
+
+# cpu_ram_efficient_loading cannot be used with MXFP4 model quantization.
+# It can only be used with a dequantized model like `axolotl-ai-co/gpt-oss-120b-dequantized`

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

@@ -53,7 +53,6 @@ warmup_ratio: 0.03
 special_tokens:
 eot_tokens:
   - "<|end|>"
-  - "<|return|>"
 
 fsdp_version: 2
 fsdp_config:

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

@@ -65,4 +65,3 @@ warmup_ratio: 0.1
 special_tokens:
 eot_tokens:
   - "<|end|>"
-  - "<|return|>"