basic torchao fp8 mixed precision training (#2926)

* debug

* debug

* debug

* revert unneeded change

* add accelerator config to base trainer builder

* add back accumulated_cache_size_limit setting

* lint

* accelerator constructor patch for single-GPU torch fp8

* lint

* re-using existing fp8 code

* lint

* remove accelerate patch now fix in latest release

* fix

* docs

* add fp8 + fsdp2 example

* remove unused config

* update config

* smoke tests

* add validator

* add 2.7.0 guard for fsdp2

* fix

* add config descriptions

* add FSDP doc link

* nit

* set force_recompute_fp8_weight_in_bwd with enable_fsdp_float8_all_gather

* better cfg for smoke tests

* add test for accelerate patching

* update fp8 validator

examples/llama-3/3b-fp8-fsdp2.yaml

@@ -0,0 +1,76 @@
+base_model: meta-llama/Llama-3.2-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+
+output_dir: ./outputs/fp8_out/
+
+sample_packing: true
+pad_to_sequence_len: true
+sequence_len: 512
+
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+torch_compile: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+num_epochs: 1
+optimizer: adamw_torch_fused
+
+cosine_constant_lr_ratio: 0
+cosine_min_lr_ratio: 1.0
+learning_rate: 2e-5
+save_only_model: true
+
+fp8: true
+fp8_enable_fsdp_float8_all_gather: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_steps: 10
+weight_decay: 0.0
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: false
+
+special_tokens:
+  pad_token: <|end_of_text|>
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config