refactor tests into separate dir

.gitignore

@@ -1,3 +1,4 @@
 float8_playground/__pycache__/*
 finetune/__pycache__/*
+tests/__pycache__/*
 tmp/*

tests/context.py

@@ -0,0 +1,5 @@
+# https://docs.python-guide.org/writing/structure/#test-suite
+
+import os
+import sys
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '../float8_playground')))

float8_playground/test.py → tests/test.py

@@ -5,6 +5,9 @@
 import torch
 import torch.nn as nn
 
+# set up float8 path
+import context
+
 from float8_utils import (
     compute_error,
     tensor_to_scale,

float8_playground/test_fsdp.sh → tests/test_fsdp.sh

@@ -1,15 +1,15 @@
 #!/bin/bash
 
 # generate the test data
-python float8_playground/test_fsdp.py --mode generate
+python tests/test_fsdp.py --mode generate
 
 # generate single GPU model output and updated state dict
-python float8_playground/test_fsdp.py --mode single_gpu
+python tests/test_fsdp.py --mode single_gpu
 
 # generate FSDP model output and updated state dict
 # the NCCL_DEBUG setting is to avoid log spew
 # the CUDA_VISIBLE_DEVICES setting is for easy debugging
-NCCL_DEBUG=WARN CUDA_VISIBLE_DEVICES=0,1 python float8_playground/test_fsdp.py --mode fsdp
+NCCL_DEBUG=WARN CUDA_VISIBLE_DEVICES=0,1 python tests/test_fsdp.py --mode fsdp
 
 # compare the outputs and state dicts and verify equivalence
-python float8_playground/test_fsdp.py --mode analyze
+python tests/test_fsdp.py --mode analyze

float8_playground/test_sam.py → tests/test_sam.py

@@ -9,6 +9,9 @@
 
 from transformers import SamModel
 
+# set up float8 path
+import context
+
 from float8_linear import swap_linear_with_float8_linear
 from float8_utils import compute_error