disable autocast for ring_flash_attn_cuda

lucidrains · Apr 6, 2024 · 5b3b22f · 5b3b22f
1 parent 5b202e4
commit 5b3b22f
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 1 deletion.
diff --git a/ring_attention_pytorch/ring_flash_attention_cuda.py b/ring_attention_pytorch/ring_flash_attention_cuda.py
@@ -7,6 +7,7 @@
 from torch import nn, einsum, Tensor
 import torch.nn.functional as F
 from torch.autograd.function import Function
+from torch.cuda.amp import autocast
 
 from ring_attention_pytorch.ring import (
     ring_pass,
@@ -481,6 +482,7 @@ class RingFlashAttentionCUDAFunction(Function):
 
     @staticmethod
     @torch.no_grad()
+    @autocast(enabled = False)
     def forward(
         ctx,
         q: Tensor,
@@ -629,6 +631,7 @@ def forward(
 
     @staticmethod
     @torch.no_grad()
+    @autocast(enabled = False)
     def backward(ctx, do):
         """ Algorithm 2 in the v2 paper """
 

diff --git a/setup.py b/setup.py
@@ -3,7 +3,7 @@
 setup(
   name = 'ring-attention-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.3.0',
+  version = '0.3.1',
   license='MIT',
   description = 'Ring Attention - Pytorch',
   author = 'Phil Wang',