vllm-project
diff --git a/‎flash_attn/cute/flash_fwd_sm100.py‎
Lines changed: 1 addition & 1 deletion b/‎flash_attn/cute/flash_fwd_sm100.py‎
Lines changed: 1 addition & 1 deletion
@@ -2029,7 +2029,7 @@ def softmax_loop(
 
             rescale_threshold = (
                 8.0 if const_expr(self.q_dtype.width == 16) else
-                4.0 if const_expr(self.q_dtype.width == 8) else
+                0.75 if const_expr(self.q_dtype.width == 8) else
                 0.0
             )
             softmax = SoftmaxSm100.create(
Original file line number	Diff line number	Diff line change
`@@ -2029,7 +2029,7 @@ def softmax_loop(`
`2029`	`2029`
`2030`	`2030`	`rescale_threshold = (`
`2031`	`2031`	`8.0 if const_expr(self.q_dtype.width == 16) else`
`2032`		`- 4.0 if const_expr(self.q_dtype.width == 8) else`
	`2032`	`+ 0.75 if const_expr(self.q_dtype.width == 8) else`
`2033`	`2033`	`0.0`
`2034`	`2034`	`)`
`2035`	`2035`	`softmax = SoftmaxSm100.create(`