diff --git a/CHANGELOG.md b/CHANGELOG.md
index fe5c4d7a0..c19e56d83 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,13 @@
 
 # CUTLASS 4.x
 
+## [4.3.6](https://github.com/NVIDIA/cutlass/releases/tag/v4.3.6) (2026-06-17)
+
+### CUTLASS C++
+* Make [version.h](https://github.com/NVIDIA/cutlass/blob/release/4.4/include/cutlass/version.h) NVRTC JIT compilation compatible.
+* Allow linking large cutlass library on 64bit platform.
+* Fix alignment-related miscalculation for pipeline stages of Blackwell blockscaled GEMM.
+
 ## [4.3.5](https://github.com/NVIDIA/cutlass/releases/tag/v4.3.5) (2026-01-09)
 
 ### CuTe DSL
diff --git a/README.md b/README.md
index af692c802..959f55c30 100644
--- a/README.md
+++ b/README.md
@@ -1,9 +1,9 @@
 ![ALT](./media/images/gemm-hierarchy-with-epilogue-no-labels.png "Complete CUDA GEMM decomposition")
 # Overview
 
-# CUTLASS 4.3.5
+# CUTLASS 4.3.6
 
-_CUTLASS 4.3.5 - Jan 2026_
+_CUTLASS 4.3.6 - June 2026_
 
 CUTLASS is a collection of abstractions for implementing high-performance matrix-matrix multiplication (GEMM)
 and related computations at all levels and scales within CUDA. It incorporates strategies for
diff --git a/include/cutlass/gemm/collective/builders/sm100_blockscaled_umma_builder.inl b/include/cutlass/gemm/collective/builders/sm100_blockscaled_umma_builder.inl
index 78b66196d..2c28c2ca4 100644
--- a/include/cutlass/gemm/collective/builders/sm100_blockscaled_umma_builder.inl
+++ b/include/cutlass/gemm/collective/builders/sm100_blockscaled_umma_builder.inl
@@ -73,7 +73,7 @@ sm100_compute_stage_count_or_override_blockscaled(StageCountAutoCarveout<carveou
   // 1. smem for A and smem for B (CollectiveMma::SharedStorage::TensorStorage)
   // 2. one MainloopPipeline = PipelineTmaUmmaAsync (CollectiveMma::SharedStorage::SharedStorage)
   // 3. smem for SFB and smem for SFB (CollectiveMma::SharedStorage::TensorStorage, independent of input size b.c. sizeof(sf) is fixed)
-  constexpr auto mainloop_pipeline_bytes = sizeof(typename cutlass::PipelineTmaUmmaAsync<1>::SharedStorage);
+  constexpr auto mainloop_pipeline_bytes = cutlass::round_up(sizeof(typename cutlass::PipelineTmaUmmaAsync<1>::SharedStorage), 128);;
   constexpr auto a_bits = cute::sizeof_bits_v<ElementA>;
   constexpr auto b_bits = cute::sizeof_bits_v<ElementB>;
   constexpr auto stage_sfa_bytes = size(filter_zeros(TileShapeSFA{}));
diff --git a/include/cutlass/version.h b/include/cutlass/version.h
index 658933e68..9f87549fd 100644
--- a/include/cutlass/version.h
+++ b/include/cutlass/version.h
@@ -31,12 +31,16 @@
 
 #pragma once
 
+#if defined(__CUDACC_RTC__)
+#include CUDA_STD_HEADER(cstdint)
+#else
 #include <cstdint>
 #include <string>
+#endif
 
 #define CUTLASS_MAJOR 4
 #define CUTLASS_MINOR 3
-#define CUTLASS_PATCH 5
+#define CUTLASS_PATCH 6
 
 #ifdef CUTLASS_VERSIONS_GENERATED
 #include "cutlass/version_extended.h"
@@ -65,6 +69,7 @@ namespace cutlass {
     return CUTLASS_BUILD + 0;
   }
 
+#if !defined(__CUDACC_RTC__)
   inline std::string getVersionString() {
     std::string version = "@CUTLASS_VERSION@";
     if (getVersionBuild()) {
@@ -72,9 +77,10 @@ namespace cutlass {
     }
     return version;
   }
-  
+
   inline std::string getGitRevision() {
     return "@CUTLASS_REVISION@";
   }
+#endif // !defined(__CUDACC_RTC__)
 
 } // namespace cutlass
diff --git a/python/cutlass_cppgen/__init__.py b/python/cutlass_cppgen/__init__.py
index 26085c3e1..fa2398f6f 100644
--- a/python/cutlass_cppgen/__init__.py
+++ b/python/cutlass_cppgen/__init__.py
@@ -133,7 +133,7 @@ def get_option_registry():
         this._option_registry = OptionRegistry(device_cc())
     return this._option_registry
 
-this.__version__ = '4.3.5'
+this.__version__ = '4.3.6'
 
 from cutlass_cppgen.backend import create_memory_pool
 from cutlass_cppgen.emit.pytorch import pytorch
diff --git a/python/setup_cutlass.py b/python/setup_cutlass.py
index d057de160..7b6a65ad6 100644
--- a/python/setup_cutlass.py
+++ b/python/setup_cutlass.py
@@ -51,7 +51,7 @@ setup_pycute.perform_setup()
 
 setup(
     name='cutlass_cppgen',
-    version='4.3.5',
+    version='4.3.6',
     description='CUTLASS Pythonic Interface',
     package_dir={'': '.'},
     packages=[
diff --git a/python/setup_library.py b/python/setup_library.py
index c8f0bc95d..f4442e5d6 100644
--- a/python/setup_library.py
+++ b/python/setup_library.py
@@ -36,7 +36,7 @@ from setuptools import setup
 def perform_setup():
     setup(
         name='cutlass_library',
-        version='4.3.5',
+        version='4.3.6',
         description='CUTLASS library generation scripts',
         packages=['cutlass_library']
     )
diff --git a/python/setup_pycute.py b/python/setup_pycute.py
index c13638782..64a703b5f 100644
--- a/python/setup_pycute.py
+++ b/python/setup_pycute.py
@@ -36,7 +36,7 @@ from setuptools import setup
 def perform_setup():
     setup(
         name='pycute',
-        version='4.3.5',
+        version='4.3.6',
         description='Python implementation of CuTe',
         packages=['pycute'],
     )
diff --git a/tools/library/CMakeLists.txt b/tools/library/CMakeLists.txt
index b512aa9c4..c8dd78ef8 100644
--- a/tools/library/CMakeLists.txt
+++ b/tools/library/CMakeLists.txt
@@ -123,6 +123,19 @@ function(cutlass_add_cutlass_library)
     PRIVATE cutlass_library_internal_interface
     )
 
+  # Use medium code model to handle large libraries (>2GB) on Linux x86_64
+  # Conservative approach: only apply on Linux with GCC/Clang/Intel compilers
+  # This addresses the "relocation truncated to fit" linker error when libraries exceed 2GB
+  if(CMAKE_SYSTEM_NAME STREQUAL "Linux" AND
+     CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64" AND
+     CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang|IntelLLVM")
+    message(STATUS "CUTLASS: Applying -mcmodel=medium for large library support (Linux x86_64, ${CMAKE_CXX_COMPILER_ID})")
+    target_compile_options(${__NAME}_objs PRIVATE
+      $<$<COMPILE_LANGUAGE:CXX>:-mcmodel=medium>
+      $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-mcmodel=medium>
+    )
+  endif()
+
   if (CUTLASS_BUILD_MONO_LIBRARY AND __SUFFIX)
 
     # If we're only building a single monolithic library then we