Activation function Triton kernels, LoRA custom autograd functions (#2324)

* LoRA + activation fn Triton kernels: initial commit

* implementing optims

* finalizing MLP LoRA kernels and progress on QKV / W kernels

* updates

* O projection optim

* adding monkey patching logic

* doc strings, typing, pre-commit fixes

* updates

* adding lora 8b kernels example

* working on fsdp support

* tests and fixes

* small fixes, getting tests to pass, adding doc strings

* integration tests for LoRA patching

* config.qmd

* remove unneeded pytest fixture

* fix

* review comments first pass

* improving tests, attention class agnostic patching

* adding support for more archs

* wip SiLU / GELU impls

* improved testing, small updates, etc.

* slightly updating docs

* rebase

* fixing test_attention_patching_integration

* additional review comments, fixing test in CI (hopefully)

* isolating problematic patching test

* relaxing allclose threshold to reduce flakiness

* fixing accidental change

* adding model arch agnostic attention class fetching

* removing unused activations

examples/llama-3/lora-1b-kernels.yml

@@ -0,0 +1,82 @@
+base_model: NousResearch/Llama-3.2-1B
+# 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
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 16
+lora_alpha: 32
+# Currently, we don't support dropout with our custom Triton kernels
+# lora_dropout: 0.05
+lora_fan_in_fan_out:
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# These options enable our custom Triton kernels / autograd
+# functions for MLP and attention calculations
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: "<|end_of_text|>"