init

Author: metascroy

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-e6f766c7d750d40603eee3f66c5915bac606b3ea
+a0a8eca01a88b01f517559fd3cf565a58c134805

runtime/core/portable_type/c10/c10/util/llvmMathExtras.h

@@ -370,7 +370,7 @@ constexpr inline bool isShiftedInt(int64_t x) {
 template <unsigned N>
 constexpr inline std::enable_if_t<(N < 64), bool> isUInt(uint64_t X) {
   static_assert(N > 0, "isUInt<0> doesn't make sense");
-  return X < (UINT64_C(1) << (N));
+  return X < (UINT64_C(1) << N);
 }
 template <unsigned N>
 constexpr inline std::enable_if_t<N >= 64, bool> isUInt(uint64_t /*X*/) {

runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h

@@ -1,6 +1,11 @@
 #ifndef C10_MACROS_MACROS_H_
 #define C10_MACROS_MACROS_H_
+
+#ifdef __cplusplus
 #include <cassert>
+#else
+#include <assert.h>
+#endif
 
 /* Main entry for torch/headeronly/macros (used to be c10/macros).
  *
@@ -139,6 +144,8 @@
 
 #define C10_RESTRICT __restrict
 
+#ifdef __cplusplus
+
 // Simply define the namespace, in case a dependent library want to refer to
 // the c10 namespace but not any nontrivial files.
 namespace c10 {}
@@ -176,6 +183,8 @@ namespace at::xpu {
 using namespace c10::xpu;
 } // namespace at::xpu
 
+#endif // __cplusplus
+
 // C10_LIKELY/C10_UNLIKELY
 //
 // These macros provide parentheses, so you can use these macros as:
@@ -236,7 +245,11 @@ using namespace c10::xpu;
 
 #define C10_ERASE C10_ALWAYS_INLINE C10_ATTR_VISIBILITY_HIDDEN
 
+#ifdef __cplusplus
 #include <cstdint>
+#else
+#include <stdint.h>
+#endif
 
 #ifdef __HIPCC__
 // Unlike CUDA, HIP requires a HIP header to be included for __host__ to work.
@@ -467,7 +480,7 @@ __host__ __device__
 // a non-negligible performance impact even if the assert condition is
 // never triggered. We choose to use abort() instead which will still
 // terminate the application but without a more useful error message.
-#if !defined(C10_USE_ROCM_KERNEL_ASSERT) and defined(USE_ROCM)
+#if !defined(C10_USE_ROCM_KERNEL_ASSERT) && defined(USE_ROCM)
 #define CUDA_KERNEL_ASSERT(cond) \
   if C10_UNLIKELY (!(cond)) {    \
     abort();                     \
@@ -517,9 +530,21 @@ __host__ __device__
     __assert_fail(                                                       \
         #cond, __FILE__, static_cast<unsigned int>(__LINE__), __func__); \
   }
-#endif //  C10_USE_ROCM_KERNEL_ASSERT and USE_ROCM
+#endif //  C10_USE_ROCM_KERNEL_ASSERT && USE_ROCM
 #endif // __APPLE__
 
+// Compile-time switch to control how assertions are logged inside CUDA kernels.
+// If C10_CUDA_VERBOSE_ASSERT is defined,  CUDA_KERNEL_ASSERT_VERBOSE will
+// take addition information passed to the macro and forward them to
+// CUDA_KERNEL_ASSERT_PRINTF If C10_CUDA_VERBOSE_ASSERT is not defined,
+// CUDA_KERNEL_ASSERT_VERBOSE will behave the same as CUDA_KERNEL_ASSERT.
+#ifdef C10_ENABLE_VERBOSE_ASSERT
+#define CUDA_KERNEL_ASSERT_VERBOSE(cond, ...) \
+  CUDA_KERNEL_ASSERT_PRINTF(cond, __VA_ARGS__)
+#else
+#define CUDA_KERNEL_ASSERT_VERBOSE(cond, ...) CUDA_KERNEL_ASSERT(cond)
+#endif
+
 #ifdef __APPLE__
 #include <TargetConditionals.h>
 #endif
@@ -611,4 +636,59 @@ __host__ __device__
 #define C10_RETURN_MOVE_IF_OLD_COMPILER 0
 #endif
 
+// The HIDDEN_NAMESPACE_BEGIN and HIDDEN_NAMESPACE_END below
+// are needed for maintaining robustness in our header APIs in
+// torch/headeronly and torch/csrc/stable under the namespaces
+// torch::headeronly and torch::stable respectively. We enforce
+// hidden visibility for these APIs because we want to enable
+// loading custom extensions compiled against different libtorch
+// versions where these APIs may have changed.
+
+// Helper macros to handle 1-3 hidden namespace levels when not windows
+#define _HIDDEN_NS_GET_MACRO(_1, _2, _3, NAME, ...) NAME
+#define _HIDDEN_NS_1(n1) namespace n1 __attribute__((visibility("hidden"))) {
+#define _HIDDEN_NS_2(n1, n2) \
+  namespace n1 {             \
+  namespace n2 __attribute__((visibility("hidden"))) {
+#define _HIDDEN_NS_3(n1, n2, n3) \
+  namespace n1::n2 {             \
+  namespace n3 __attribute__((visibility("hidden"))) {
+
+// Helper macros to close namespaces when not windows
+#define _HIDDEN_NS_END_1(n1) }
+#define _HIDDEN_NS_END_N(n1, ...) \
+  }                               \
+  }
+
+// Helper macros to join strs with :: (for win, where symbols are hidden by
+// default)
+#define _EXPAND(...) __VA_ARGS__
+#define _JOIN_GET_MACRO(_1, _2, _3, NAME, ...) NAME
+#define _JOIN_NS1(a) a
+#define _JOIN_NS2(a, b) a::b
+#define _JOIN_NS3(a, b, c) a::b::c
+
+#if !defined(HIDDEN_NAMESPACE_BEGIN)
+#if defined(__GNUG__) && !defined(_WIN32)
+#define HIDDEN_NAMESPACE_BEGIN(...) \
+  _HIDDEN_NS_GET_MACRO(             \
+      __VA_ARGS__, _HIDDEN_NS_3, _HIDDEN_NS_2, _HIDDEN_NS_1)(__VA_ARGS__)
+#else
+#define HIDDEN_NAMESPACE_BEGIN(...)  \
+  namespace _EXPAND(_JOIN_GET_MACRO( \
+      __VA_ARGS__, _JOIN_NS3, _JOIN_NS2, _JOIN_NS1)(__VA_ARGS__)) {
+#endif
+#endif
+
+#if !defined(HIDDEN_NAMESPACE_END)
+#if defined(__GNUG__) && !defined(_WIN32)
+#define HIDDEN_NAMESPACE_END(...)                                         \
+  _HIDDEN_NS_GET_MACRO(                                                   \
+      __VA_ARGS__, _HIDDEN_NS_END_N, _HIDDEN_NS_END_N, _HIDDEN_NS_END_1)( \
+      __VA_ARGS__)
+#else
+#define HIDDEN_NAMESPACE_END(...) }
+#endif
+#endif
+
 #endif // C10_MACROS_MACROS_H_

runtime/core/portable_type/c10/torch/headeronly/util/BFloat16.h

@@ -395,7 +395,7 @@ inline C10_HOST_DEVICE bool operator<(BFloat16& lhs, BFloat16& rhs) {
 C10_CLANG_DIAGNOSTIC_POP()
 } // namespace c10
 
-namespace torch::headeronly {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly)
 
 namespace detail {
 using c10::detail::bits_from_f32;
@@ -415,7 +415,7 @@ using c10::operator/=;
 using c10::operator<;
 using c10::operator>;
 using c10::operator<<;
-} // namespace torch::headeronly
+HIDDEN_NAMESPACE_END(torch, headeronly)
 
 namespace std {
 

runtime/core/portable_type/c10/torch/headeronly/util/Half.h

@@ -698,7 +698,7 @@ C10_CLANG_DIAGNOSTIC_POP()
 
 } // namespace c10
 
-namespace torch::headeronly {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly)
 
 using c10::Half;
 using c10::operator+;
@@ -724,7 +724,7 @@ using c10::detail::fp16_ieee_to_fp32_bits;
 using c10::detail::fp16_ieee_to_fp32_value;
 } // namespace detail
 
-} // namespace torch::headeronly
+HIDDEN_NAMESPACE_END(torch, headeronly)
 
 namespace std {
 

runtime/core/portable_type/c10/torch/headeronly/util/TypeSafeSignMath.h

@@ -79,7 +79,7 @@ template <typename Limit, typename T>
 inline constexpr bool greater_than_max(const T& x) {
   constexpr bool can_overflow =
       std::numeric_limits<T>::digits > std::numeric_limits<Limit>::digits;
-  return can_overflow && x > (std::numeric_limits<Limit>::max)();
+  return can_overflow && x > std::numeric_limits<Limit>::max();
 }
 
 #ifdef __GNUC__
@@ -139,10 +139,10 @@ inline constexpr bool less_than_lowest(const T& x) {
 
 C10_CLANG_DIAGNOSTIC_POP()
 
-namespace torch::headeronly {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly)
 using c10::greater_than_max;
 using c10::is_negative;
 using c10::less_than_lowest;
 using c10::signs_differ;
 using c10::signum;
-} // namespace torch::headeronly
+HIDDEN_NAMESPACE_END(torch, headeronly)

runtime/core/portable_type/c10/torch/headeronly/util/bit_cast.h

@@ -13,7 +13,7 @@
 #endif // __has_include(<bit>) && (__cplusplus >= 202002L ||
        // (defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L))
 
-namespace torch::headeronly {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly)
 
 #if C10_HAVE_STD_BIT_CAST
 using std::bit_cast;
@@ -43,7 +43,7 @@ bit_cast(const From& src) noexcept {
 #endif // C10_HAVE_STD_BIT_CAST
 #undef C10_HAVE_STD_BIT_CAST
 
-} // namespace torch::headeronly
+HIDDEN_NAMESPACE_END(torch, headeronly)
 
 namespace c10 {
 using torch::headeronly::bit_cast;

runtime/core/portable_type/c10/torch/headeronly/util/complex.h

@@ -590,7 +590,7 @@ struct alignas(4) complex<Half> {
 
 } // namespace c10
 
-namespace torch::headeronly {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly)
 using c10::complex;
 using c10::operator+;
 using c10::operator-;
@@ -611,6 +611,6 @@ using c10::complex_literals::operator""_if;
 using c10::complex_literals::operator""_id;
 } // namespace complex_literals
 
-} // namespace torch::headeronly
+HIDDEN_NAMESPACE_END(torch, headeronly)
 
 C10_CLANG_DIAGNOSTIC_POP()

runtime/core/portable_type/c10/torch/headeronly/util/floating_point_utils.h

@@ -4,7 +4,7 @@
 #include <torch/headeronly/util/bit_cast.h>
 #include <cstdint>
 
-namespace torch::headeronly::detail {
+HIDDEN_NAMESPACE_BEGIN(torch, headeronly, detail)
 
 C10_HOST_DEVICE inline float fp32_from_bits(uint32_t w) {
 #if defined(__OPENCL_VERSION__)
@@ -30,7 +30,7 @@ C10_HOST_DEVICE inline uint32_t fp32_to_bits(float f) {
 #endif
 }
 
-} // namespace torch::headeronly::detail
+HIDDEN_NAMESPACE_END(torch, headeronly, detail)
 
 namespace c10::detail {
 using torch::headeronly::detail::fp32_from_bits;

torch_pin.py

@@ -1,2 +1,2 @@
 TORCH_VERSION = "2.10.0"
-NIGHTLY_VERSION = "dev20251015"
+NIGHTLY_VERSION = "dev20251104"