make phi training work with Loras

examples/phi/README.md

@@ -1,7 +1,11 @@
 # Phi
 
-Due to some nuances with the phi code, please use deepspeed when training phi.
+Due to some nuances with the phi code, please use deepspeed when training phi for full finetune.
 
 ```shell
-accelerate launch scripts/finetune.py examples/phi/phi-ft.yml --deepspeed deepspeed/zero1.json
+accelerate launch -m axolotl.cli.train examples/phi/phi-ft.yml --deepspeed deepspeed/zero1.json
+
+# OR
+
+python -m axolotl.cli.train examples/phi/phi-qlora.yml
 ```

examples/phi/phi-qlora.yml

@@ -0,0 +1,75 @@
+base_model: microsoft/phi-1_5
+base_model_config: microsoft/phi-1_5
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+is_llama_derived_model: false
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: garage-bAInd/Open-Platypus
+    type: alpaca
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./phi-sft-out
+
+sequence_len: 1024
+sample_packing: false  # not CURRENTLY compatible with LoRAs
+pad_to_sequence_len:
+
+adapter: qlora
+lora_model_dir:
+lora_r: 64
+lora_alpha: 32
+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: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_torch
+adam_beta2: 0.95
+adam_epsilon: 0.00001
+max_grad_norm: 1.0
+lr_scheduler: cosine
+learning_rate: 0.000003
+
+train_on_inputs: false
+group_by_length: true
+bf16: true
+fp16: false
+tf32: true
+
+gradient_checkpointing:
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+
+warmup_steps: 100
+eval_steps: 0.05
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.1
+fsdp:
+fsdp_config:
+resize_token_embeddings_to_32x: true
+special_tokens:
+  bos_token: "<|endoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<|endoftext|>"
+  pad_token: "<|endoftext|>"

src/axolotl/utils/config.py

@@ -75,6 +75,7 @@ def normalize_config(cfg):
         cfg.torch_dtype = torch.float32
 
     model_config = load_model_config(cfg)
+    cfg.model_config_type = model_config.model_type
 
     # figure out if the model is llama
     cfg.is_llama_derived_model = (
@@ -237,6 +238,21 @@ def validate_config(cfg):
             raise ValueError(
                 "`early_stopping_patience` requires that eval_steps should evenly divide save_steps."
             )
+
+    if cfg.model_type == "MixFormerSequentialForCausalLM" and cfg.adapter is not None:
+        LOG.warning("Use AutoModelForCausalLM for phi/MixFormer models with qLoRA")
+
+    if cfg.model_config_type == "mixformer-sequential":
+        if cfg.sample_packing:
+            if cfg.adapter is not None:
+                LOG.warning(
+                    "phi/MixFormer models are not currently compatible with LoRA and sample_packing"
+                )
+            if cfg.model_type == "AutoModelForCausalLM":
+                raise ValueError(
+                    "`model_type: MixFormerSequentialForCausalLM` required for sample_packing"
+                )
+
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/models.py

@@ -1,6 +1,5 @@
 """Module for models and model loading"""
-
-
+import importlib
 import logging
 import math
 import os
@@ -155,11 +154,26 @@ def load_model(
         LOG.info("patching _expand_mask")
         hijack_expand_mask()
 
+    model_config = load_model_config(cfg)
+
+    # special handling b/c remote MixFormers code doesn't have _no_split_modules set
+    if (
+        "MixFormerSequentialConfig" in model_config.__class__.__name__
+        and cfg.model_type == "AutoModelForCausalLM"
+    ):
+        module_name = model_config.__class__.__module__.replace(
+            ".configuration_mixformer_sequential", ".modeling_mixformer_sequential"
+        )
+        modeling_phi = importlib.import_module(module_name)
+        # pylint:disable=protected-access
+        modeling_phi.MixFormerSequentialForCausalLM._no_split_modules = [
+            "ParallelBlock"
+        ]
+
     model_kwargs = {}
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
     if cfg.gptq:
-        model_config = load_model_config(cfg)
         if not hasattr(model_config, "quantization_config"):
             LOG.warning("model config does not contain quantization_config information")
         else: