refactor inference, warn if model is frozen

scripts/finetune.py

@@ -6,9 +6,11 @@ import random
 import signal
 import sys
 from pathlib import Path
+from typing import Optional
 
 import fire
 import torch
+import transformers
 import yaml
 from attrdict import AttrDefault
 
@@ -46,6 +48,15 @@ def choose_device(cfg):
         cfg.device_map = {"": cfg.device}
 
 
+def get_multi_line_input() -> Optional[str]:
+    print("Give me an instruction (Ctrl + Z to finish): ")
+    instruction = ""
+    for line in sys.stdin:
+        instruction += line
+    # instruction = pathlib.Path("/proc/self/fd/0").read_text()
+    return instruction
+
+
 def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
     tokenizer.add_special_tokens({"unk_token": "<unk>"})
     tokenizer.add_special_tokens({"bos_token": "<s>"})
@@ -55,8 +66,7 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
 
     while True:
         # support for multiline inputs
-        print("Give me an instruction (Ctrl + D to finish): ")
-        instruction = pathlib.Path("/proc/self/fd/0").read_text()
+        instruction = get_multi_line_input()
         if not instruction:
             return
         prompt = prompter_module().build_prompt(instruction=instruction)
@@ -66,7 +76,7 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
         with torch.no_grad():
             # gc = GenerationConfig()  # TODO swap out and use this
             generated = model.generate(
-                inputs=batch["input_ids"].to("cuda"),
+                inputs=batch["input_ids"].to(cfg.device),
                 do_sample=True,
                 use_cache=True,
                 repetition_penalty=1.1,

src/axolotl/utils/models.py

@@ -183,6 +183,12 @@ def load_model(
         model.is_parallelizable = True
         model.model_parallel = True
 
+    requires_grad = []
+    for name, param in model.named_parameters(recurse=True):
+        if param.requires_grad:
+            requires_grad.append(f"{name}: {param.requires_grad}")
+    if len(requires_grad) == 0:
+        logging.warning("there are no parameters that require gradient updates")
 
     # TODO resume_from_checkpoint handling
     return model, tokenizer, lora_config

src/axolotl/utils/trainer.py

@@ -105,7 +105,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         run_name=cfg.wandb_run_id if cfg.use_wandb else None,
         optim=cfg.optimizer if cfg.optimizer else None,
         lr_scheduler_type=cfg.lr_scheduler if cfg.lr_scheduler not in ("one_cycle", "log_sweep") else "cosine",
-        weight_decay=cfg.weight_decay if cfg.weight_decay else 0.0,
+        weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
         **training_arguments_kwargs,
     )