diff --git a/transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py b/transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py
index cca013461d..0b3dd1a0e1 100644
--- a/transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py
+++ b/transformer_engine/pytorch/attention/dot_product_attention/context_parallel.py
@@ -27,7 +27,7 @@
 from transformer_engine.pytorch.tensor.float8_tensor import Float8Tensor
 from transformer_engine.pytorch.tensor.storage.float8_tensor_storage import Float8TensorStorage
 from transformer_engine.pytorch.quantized_tensor import QuantizedTensorStorage
-from transformer_engine.pytorch.jit import jit_fuser
+from transformer_engine.pytorch.jit import jit_fuser, jit_fuser_dynamic
 from transformer_engine.pytorch.graph import is_graph_capturing
 from transformer_engine.pytorch.constants import dist_group_type
 from transformer_engine.pytorch.distributed import (
@@ -184,7 +184,7 @@ def flash_attn_fwd_second_half_out_correction(
     out_.add_(out_corrected)
 
 
-@jit_fuser
+@jit_fuser_dynamic
 def flash_attn_fwd_softmax_lse_correction(
     softmax_lse: torch.Tensor,
     softmax_lse_per_step: torch.Tensor,
@@ -196,7 +196,7 @@ def flash_attn_fwd_softmax_lse_correction(
     softmax_lse.copy_(new_scale)
 
 
-@jit_fuser
+@jit_fuser_dynamic
 def flash_attn_fwd_second_half_softmax_lse_correction(
     softmax_lse: torch.Tensor,
     softmax_lse_per_step: torch.Tensor,
diff --git a/transformer_engine/pytorch/jit.py b/transformer_engine/pytorch/jit.py
index 4d52d9b925..25089d3bbd 100644
--- a/transformer_engine/pytorch/jit.py
+++ b/transformer_engine/pytorch/jit.py
@@ -36,9 +36,30 @@ def wrapper(*args, **kwargs):
     return wrapper
 
 
+# See: https://github.com/ROCm/TransformerEngine/issues/693
+def lazy_compile_dynamic(func):
+    """Lazy compile a function with torch.compile, using dynamic shapes
+
+    Compiling dynamic up front skips the static specialization dynamo would otherwise build on
+    the first shape it sees, leaving one compiled artifact for the process instead of two.
+    """
+    compiled_func = None
+
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        nonlocal compiled_func
+        if compiled_func is None:
+            compiled_func = torch.compile(func, dynamic=True)
+        return compiled_func(*args, **kwargs)
+
+    return wrapper
+
+
 jit_fuser = lambda func: func
+jit_fuser_dynamic = lambda func: func
 if torch_version() >= (2, 0, 0) and bool(int(os.getenv("NVTE_TORCH_COMPILE", "1"))):
     jit_fuser = lazy_compile
+    jit_fuser_dynamic = lazy_compile_dynamic
 
 
 # See: https://github.com/NVIDIA/TransformerEngine/issues/597