pin bump with better architecture

Pull Request resolved: #15016


This diff bump torch pin to 1004.

also expose cuda version into torch_pin.py for better update if needed;
ghstack-source-id: 315575273
@exported-using-ghexport

Differential Revision: [D84280496](https://our.internmc.facebook.com/intern/diff/D84280496/)

Author: Gasoonjia

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-53a2908a10f414a2f85caa06703a26a40e873869
+cf9d09490c7f6685ec68d5db3acf2e0d73c54d00

docs/source/conf.py

@@ -24,7 +24,7 @@
 import sys
 from typing import Any
 
-import pytorch_sphinx_theme2  # type: ignore[import-untyped]
+import pytorch_sphinx_theme2  # type: ignore[import-not-found]
 
 # To let us import ./custom_directives.py
 sys.path.insert(0, os.path.abspath("."))

install_requirements.py

@@ -12,33 +12,12 @@
 
 from install_utils import determine_torch_url, is_intel_mac_os, python_is_compatible
 
-from torch_pin import NIGHTLY_VERSION, TORCH_VERSION
+from torch_pin import NIGHTLY_VERSION, SUPPORTED_CUDA_VERSIONS, TORCH_VERSION
 
 # The pip repository that hosts nightly torch packages.
 # This will be dynamically set based on CUDA availability and CUDA backend enabled/disabled.
 TORCH_NIGHTLY_URL_BASE = "https://download.pytorch.org/whl/nightly"
 
-# Supported CUDA versions - modify this to add/remove supported versions
-# Format: tuple of (major, minor) version numbers
-SUPPORTED_CUDA_VERSIONS = (
-    (12, 6),
-    (12, 8),
-    (13, 0),
-)
-
-# Since ExecuTorch often uses main-branch features of pytorch, only the nightly
-# pip versions will have the required features.
-#
-# NOTE: If a newly-fetched version of the executorch repo changes the value of
-# NIGHTLY_VERSION, you should re-run this script to install the necessary
-# package versions.
-#
-# NOTE: If you're changing, make the corresponding change in .ci/docker/ci_commit_pins/pytorch.txt
-# by picking the hash from the same date in https://hud.pytorch.org/hud/pytorch/pytorch/nightly/
-#
-# NOTE: If you're changing, make the corresponding supported CUDA versions in
-# SUPPORTED_CUDA_VERSIONS above if needed.
-
 
 def install_requirements(use_pytorch_nightly):
     # Skip pip install on Intel macOS if using nightly.

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

@@ -359,6 +359,7 @@ static inline int C10_WARP_SIZE_INTERNAL() {
 // Those platforms do not support assert()
 #define CUDA_KERNEL_ASSERT(cond)
 #define CUDA_KERNEL_ASSERT_MSG(cond, msg)
+#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...)
 #define SYCL_KERNEL_ASSERT(cond)
 #elif defined(_MSC_VER)
 #if defined(NDEBUG)
@@ -396,6 +397,26 @@ __host__ __device__
                static_cast<unsigned>(__LINE__)), \
            0);                                   \
   }
+#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...)                     \
+  if (C10_UNLIKELY(!(cond))) {                                        \
+    (void)(printf(                                                    \
+        "[CUDA_KERNEL_ASSERT] " __FILE__ ":" C10_STRINGIZE(           \
+            __LINE__) ": %s: block: [%d,%d,%d], thread: [%d,%d,%d]: " \
+                      "Assertion failed: `" #cond "`: " msg "\n",     \
+        __func__,                                                     \
+        blockIdx.x,                                                   \
+        blockIdx.y,                                                   \
+        blockIdx.z,                                                   \
+        threadIdx.x,                                                  \
+        threadIdx.y,                                                  \
+        threadIdx.z,                                                  \
+        ##__VA_ARGS__));                                              \
+    (void)(_wassert(                                                  \
+               _CRT_WIDE(#cond),                                      \
+               _CRT_WIDE(__FILE__),                                   \
+               static_cast<unsigned>(__LINE__)),                      \
+           0);                                                        \
+  }
 #define SYCL_KERNEL_ASSERT(cond)                 \
   if (C10_UNLIKELY(!(cond))) {                   \
     (void)(_wassert(                             \
@@ -455,6 +476,10 @@ __host__ __device__
   if C10_UNLIKELY (!(cond)) {             \
     abort();                              \
   }
+#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...) \
+  if C10_UNLIKELY (!(cond)) {                     \
+    abort();                                      \
+  }
 #define SYCL_KERNEL_ASSERT(cond) \
   if C10_UNLIKELY (!(cond)) {    \
     abort();                     \
@@ -470,6 +495,23 @@ __host__ __device__
     __assert_fail(                                                     \
         msg, __FILE__, static_cast<unsigned int>(__LINE__), __func__); \
   }
+#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...)                        \
+  if (C10_UNLIKELY(!(cond))) {                                           \
+    printf(                                                            \
+        "[CUDA_KERNEL_ASSERT] " __FILE__ ":" C10_STRINGIZE(            \
+            __LINE__) ": %s: block: [%d,%d,%d], thread: [%d,%d,%d]: "  \
+            "Assertion failed: `" #cond "`: " msg "\n",                \
+        __func__,                                                      \
+        blockIdx.x,                                                    \
+        blockIdx.y,                                                    \
+        blockIdx.z,                                                    \
+        threadIdx.x,                                                   \
+        threadIdx.y,                                                   \
+        threadIdx.z,                                                   \
+        ##__VA_ARGS__); \
+    __assert_fail(                                                       \
+        #cond, __FILE__, static_cast<unsigned int>(__LINE__), __func__); \
+  }
 #define SYCL_KERNEL_ASSERT(cond)                                         \
   if (C10_UNLIKELY(!(cond))) {                                           \
     __assert_fail(                                                       \

torch_pin.py

@@ -1,2 +1,19 @@
+# Since ExecuTorch often uses main-branch features of pytorch, only the nightly
+# pip versions will have the required features.
+#
+# NOTE: If a newly-fetched version of the executorch repo changes the value of
+# NIGHTLY_VERSION, you should re-run install_executorch.sh script to install the necessary
+# package versions.
+#
+# NOTE: If you're changing, make the corresponding change in .ci/docker/ci_commit_pins/pytorch.txt
+# by picking the hash from the same date in https://hud.pytorch.org/hud/pytorch/pytorch/nightly/
+#
+# NOTE: If you're changing, make the corresponding supported CUDA versions in
+# SUPPORTED_CUDA_VERSIONS above if needed.
 TORCH_VERSION = "2.10.0"
-NIGHTLY_VERSION = "dev20251003"
+NIGHTLY_VERSION = "dev20251004"
+SUPPORTED_CUDA_VERSIONS = (
+    (12, 6),
+    (12, 8),
+    (13, 0),
+)