USN 0.2.0__tar.gz → 0.3.2__tar.gz

{usn-0.2.0/USN.egg-info → usn-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: USN
-Version: 0.2.0
+Version: 0.3.2
 Summary: Unified State Network (USN) Architecture Library - A production-grade autoregressive sequence modeling architecture with O(n) training and O(1) inference.
 Author: BUEORM
 License-Expression: MIT

{usn-0.2.0 → usn-0.3.2/USN.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: USN
-Version: 0.2.0
+Version: 0.3.2
 Summary: Unified State Network (USN) Architecture Library - A production-grade autoregressive sequence modeling architecture with O(n) training and O(1) inference.
 Author: BUEORM
 License-Expression: MIT

{usn-0.2.0 → usn-0.3.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "USN"
-version = "0.2.0"
+version = "0.3.2"
 description = "Unified State Network (USN) Architecture Library - A production-grade autoregressive sequence modeling architecture with O(n) training and O(1) inference."
 readme = "README.md"
 license = "MIT"

{usn-0.2.0 → usn-0.3.2}/usn/layers/block.py

@@ -71,16 +71,11 @@ def _recurrence_loop(
     return all_s, all_R
 
 
-# Try to compile the recurrence for massive speedup on GPU
-# Only compile on CUDA — on CPU the eager loop is fine
+# torch.compile is disabled by default because the sequential recurrence
+# loop with DataParallel causes issues with CUDA Graphs and inductor.
+# The recurrence is inherently sequential (g_t depends on S_{t-1}).
+# For GPU speedup, use shorter seq_len or a custom Triton kernel.
 _compiled_recurrence = _recurrence_loop
-if torch.cuda.is_available():
-    try:
-        _compiled_recurrence = torch.compile(
-            _recurrence_loop, mode="reduce-overhead", fullgraph=False
-        )
-    except Exception:
-        _compiled_recurrence = _recurrence_loop
 
 
 class USNBlock(nn.Module):