Tidy up project structure

Author: lantiga

evaluate.py

@@ -10,6 +10,10 @@
 import torch
 import tqdm
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from lit_llama import Tokenizer
 from lit_llama.adapter import LLaMA
 from lit_llama.utils import EmptyInitOnDevice, lazy_load, llama_model_lookup

evaluate_adapter.py renamed to evaluate/adapter.py

@@ -10,6 +10,10 @@
 import torch
 import tqdm
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from lit_llama import LLaMA, Tokenizer
 from lit_llama.utils import EmptyInitOnDevice
 

evaluate_full.py renamed to evaluate/full.py

@@ -10,6 +10,10 @@
 import torch
 import tqdm
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from lit_llama import LLaMA, Tokenizer
 from lit_llama.utils import EmptyInitOnDevice, lazy_load, llama_model_lookup
 from lit_llama.lora import lora

evaluate_lora.py renamed to evaluate/lora.py

@@ -12,6 +12,7 @@
 `torch.backends.cuda.enable_flash_sdp(False)` in the script below (see https://github.com/Lightning-AI/lit-llama/issues/101).
 """
 import os
+import sys
 import time
 from pathlib import Path
 import shutil
@@ -20,6 +21,10 @@
 import numpy as np
 import torch
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from generate import generate
 from lit_llama.adapter import LLaMA, LLaMAConfig, mark_only_adapter_as_trainable, adapter_state_from_state_dict
 from lit_llama.tokenizer import Tokenizer

finetune_adapter.py renamed to finetune/adapter.py

@@ -4,6 +4,8 @@
 Note: If you run into a CUDA error "Expected is_sm80 to be true, but got false", uncomment the line
 `torch.backends.cuda.enable_flash_sdp(False)` in the script below (see https://github.com/Lightning-AI/lit-llama/issues/101).
 """
+import sys
+from pathlib import Path
 import os
 import time
 from functools import partial
@@ -14,6 +16,10 @@
 import torch
 from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from generate import generate
 from lit_llama.model import Block, LLaMA, LLaMAConfig
 from lit_llama.tokenizer import Tokenizer

finetune_full.py renamed to finetune/full.py

@@ -4,13 +4,19 @@
 Note: If you run into a CUDA error "Expected is_sm80 to be true, but got false", uncomment the line
 `torch.backends.cuda.enable_flash_sdp(False)` in the script below (see https://github.com/Lightning-AI/lit-llama/issues/101).
 """
+import sys
+from pathlib import Path
 import os
 import time
 
 import lightning as L
 import numpy as np
 import torch
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from generate import generate
 from lit_llama.lora import mark_only_lora_as_trainable, lora, lora_state_dict
 from lit_llama.model import LLaMA, LLaMAConfig

finetune_lora.py renamed to finetune/lora.py

@@ -7,6 +7,10 @@
 import lightning as L
 import torch
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from lit_llama import LLaMA, Tokenizer
 from lit_llama.utils import EmptyInitOnDevice, lazy_load, llama_model_lookup
 

generate.py

@@ -7,6 +7,10 @@
 import lightning as L
 import torch
 
+# support running without installing as a package
+wd = Path(__file__).parent.parent.resolve()
+sys.path.append(str(wd))
+
 from generate import generate
 from lit_llama import Tokenizer
 from lit_llama.adapter import LLaMA