Fix seq min length issue in kernel test

environment.yml

@@ -28,6 +28,6 @@ dependencies:
       - wandb==0.12.21
       - modelcif==0.7
       - git+https://github.com/NVIDIA/dllogger.git
-      - git+https://github.com/microsoft/DeepSpeed.git
+      - git+https://github.com/cctry/DeepSpeed.git
       # TODO: Replace above when version becomes available
       # - deepspeed==0.10.4

tests/config.py

@@ -3,7 +3,7 @@ import ml_collections as mlc
 consts = mlc.ConfigDict(
     {
         "batch_size": 2,
-        "n_res": 20,
+        "n_res": 11,
         "n_seq": 13,
         "n_templ": 3,
         "n_extra": 17,

tests/test_deepspeed_evo_attention.py

@@ -27,14 +27,8 @@ from openfold.model.primitives import (
 )
 from tests.config import consts
 import tests.compare_utils as compare_utils
-from tests.data_utils import (
-    random_template_feats,
-    random_extra_msa_feats,
-)
 
-from openfold.config import model_config
 from openfold.data import data_transforms
-from openfold.model.model import AlphaFold
 from openfold.utils.tensor_utils import tensor_tree_map
 
 
@@ -70,8 +64,8 @@ class TestDeepSpeedKernel(unittest.TestCase):
         Set dtype to confirm the kernel can be used during both training (BF16) and inference (FP32),
         since the kernel itself can run with either BF16 or FP16 precision.
         """
-        n_res = consts.n_res
-        n_seq = consts.n_seq
+        n_res = 20
+        n_seq = 18
         eps = 2e-2
 
         activations = {
@@ -156,6 +150,7 @@ class TestDeepSpeedKernel(unittest.TestCase):
         batch = tensor_tree_map(move_dim, batch)
 
         with torch.no_grad():
+            with torch.cuda.amp.autocast(dtype=torch.bfloat16):
                 model = compare_utils.get_global_pretrained_openfold()
                 out_repro = model(batch)
 
@@ -169,7 +164,8 @@ class TestDeepSpeedKernel(unittest.TestCase):
                 out_repro = out_repro["sm"]["positions"][-1].squeeze(0)
                 out_repro_ds = out_repro_ds["sm"]["positions"][-1].squeeze(0)
 
-            self.assertTrue(torch.max(torch.abs(out_repro - out_repro_ds)) < eps)
+                err = torch.max(torch.abs(out_repro - out_repro_ds))
+                self.assertTrue(err < eps, f'Error: {err}')
 
 
 if __name__ == "__main__":