[Inductor][Grouped Gemm] Fix Logic Enabling CuTeDSL Grouped Gemm

Summary: Fixes test failure in T243876826

Test Plan: `buck test fbcode//mode/opt -c 'fbcode.nvcc_arch=h100a' fbcode//pytorch/tritonbench/test/test_gpu:test_gpu -- --exact 'fbcode//pytorch/tritonbench/test/test_gpu:test_gpu - test_gpu_tritonbench_grouped_gemm (pytorch.tritonbench.test.test_gpu.main.TestTritonbenchGpu)'`

Differential Revision: D86208205

.ci/docker/requirements-docs.txt

@@ -1,11 +1,15 @@
-sphinx==7.2.6
+sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
-#Pinned versions: 7.2.6
+#Pinned versions: 5.3.0
 
-pytorch_sphinx_theme2==0.2.0
-#Description: This is needed to generate PyTorch docs
-#Pinned versions: 0.2.0
+standard-imghdr==3.13.0; python_version >= "3.13"
+#Description: This is needed by Sphinx, so it needs to be added here.
+# The reasons are as follows:
+# 1) This module has been removed from the Python standard library since Python 3.13(https://peps.python.org/pep-0594/#imghdr);
+# 2) The current version of Sphinx (5.3.0) is not compatible with Python 3.13.
+# Once Sphinx is upgraded to a version compatible with Python 3.13 or later, we can remove this dependency.
 
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@71e55749be14ceb56e7f8211a9fb649866b87ad4#egg=pytorch_sphinx_theme2
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
 # but it doesn't seem to work and hangs around idly. The initial thought that it is probably
 # something related to Docker setup. We can investigate this later.
@@ -32,17 +36,17 @@ tensorboard==2.18.0 ; python_version >= "3.13"
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 2.13.0
 
-breathe==4.36.0
+breathe==4.34.0
 #Description: This is used to generate PyTorch C++ docs
-#Pinned versions: 4.36.0
+#Pinned versions: 4.34.0
 
-exhale==0.3.7
+exhale==0.2.3
 #Description: This is used to generate PyTorch C++ docs
-#Pinned versions: 0.3.7
+#Pinned versions: 0.2.3
 
-docutils==0.20
+docutils==0.16
 #Description: This is used to generate PyTorch C++ docs
-#Pinned versions: 0.20
+#Pinned versions: 0.16
 
 bs4==0.0.1
 #Description: This is used to generate PyTorch C++ docs
@@ -52,13 +56,13 @@ IPython==8.12.0
 #Description: This is used to generate PyTorch functorch docs
 #Pinned versions: 8.12.0
 
-myst-nb==1.3.0
+myst-nb==0.17.2
 #Description: This is used to generate PyTorch functorch and torch.compile docs.
-#Pinned versions: 1.3.0
+#Pinned versions: 0.17.2
 
 # The following are required to build torch.distributed.elastic.rendezvous.etcd* docs
 python-etcd==0.4.5
 sphinx-copybutton==0.5.0
-sphinx-design==0.6.1
+sphinx-design==0.4.0
 sphinxcontrib-mermaid==1.0.0
-myst-parser==4.0.1
+myst-parser==0.18.1

.ci/pytorch/python_doc_push_script.sh

@@ -89,41 +89,23 @@ if [ "$is_main_doc" = true ]; then
 
   make coverage
   # Now we have the coverage report, we need to make sure it is empty.
-  # Sphinx 7.2.6+ format: python.txt contains a statistics table with a TOTAL row
-  # showing the undocumented count in the third column.
-  # Example: | TOTAL | 99.83% | 2 |
+  # Count the number of lines in the file and turn that number into a variable
+  # $lines. The `cut -f1 ...` is to only parse the number, not the filename
+  # Skip the report header by subtracting 2: the header will be output even if
+  # there are no undocumented items.
   #
   # Also: see docs/source/conf.py for "coverage_ignore*" items, which should
   # be documented then removed from there.
-
-  # Extract undocumented count from TOTAL row in Sphinx 7.2.6 statistics table
-  # The table format is: | Module | Coverage | Undocumented |
-  # Extract the third column (undocumented count) from the TOTAL row
-  undocumented=$(grep "| TOTAL" build/coverage/python.txt | awk -F'|' '{print $4}' | tr -d ' ')
-
-  if [ -z "$undocumented" ] || ! [[ "$undocumented" =~ ^[0-9]+$ ]]; then
+  lines=$(wc -l build/coverage/python.txt 2>/dev/null |cut -f1 -d' ')
+  undocumented=$((lines - 2))
+  if [ $undocumented -lt 0 ]; then
     echo coverage output not found
     exit 1
-  elif [ "$undocumented" -gt 0 ]; then
-    set +x  # Disable command echoing for cleaner output
-    echo ""
-    echo "====================="
-    echo "UNDOCUMENTED OBJECTS:"
-    echo "====================="
-    echo ""
-    # Find the line number of the TOTAL row and print only what comes after it
-    total_line=$(grep -n "| TOTAL" build/coverage/python.txt | cut -d: -f1)
-    if [ -n "$total_line" ]; then
-      # Print only the detailed list (skip the statistics table)
-      tail -n +$((total_line + 2)) build/coverage/python.txt
-    else
-      # Fallback to showing entire file if TOTAL line not found
-      cat build/coverage/python.txt
-    fi
-    echo ""
+  elif [ $undocumented -gt 0 ]; then
+    echo undocumented objects found:
+    cat build/coverage/python.txt
     echo "Make sure you've updated relevant .rsts in docs/source!"
-    echo "You can reproduce locally by running 'cd docs && make coverage && tail -n +\$((grep -n \"| TOTAL\" build/coverage/python.txt | cut -d: -f1) + 2)) build/coverage/python.txt'"
-    set -x  # Re-enable command echoing
+    echo "You can reproduce locally by running 'cd docs && make coverage && cat build/coverage/python.txt'"
     exit 1
   fi
 else

.github/workflows/inductor-unittest.yml

@@ -115,10 +115,10 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
         { include: [
-          { config: "inductor_amx", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "inductor_amx", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "inductor_avx2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.avx2" },
-          { config: "inductor_avx2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.avx2" },
+          { config: "inductor_amx", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "inductor_amx", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "inductor_avx2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.10xlarge.avx2" },
+          { config: "inductor_avx2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.10xlarge.avx2" },
         ]}
     secrets: inherit
 

.github/workflows/inductor.yml

@@ -84,13 +84,13 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       test-matrix: |
         { include: [
-          { config: "cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_huggingface", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_timm", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_timm", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
+          { config: "cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_huggingface", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_timm", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_timm", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
           { config: "inductor_torchbench_cpu_smoketest_perf", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.24xl.spr-metal" },
         ]}
       build-additional-packages: "vision audio torchao"

SECURITY.md

@@ -1,7 +1,7 @@
 # Security Policy
 
  - [**Reporting a Vulnerability**](#reporting-a-vulnerability)
- - [**Using PyTorch Securely**](#using-pytorch-securely)
+ - [**Using Pytorch Securely**](#using-pytorch-securely)
    - [Untrusted models](#untrusted-models)
    - [TorchScript models](#torchscript-models)
    - [Untrusted inputs](#untrusted-inputs)
@@ -10,28 +10,28 @@
 - [**CI/CD security principles**](#cicd-security-principles)
 ## Reporting Security Issues
 
-Beware that none of the topics under [Using PyTorch Securely](#using-pytorch-securely) are considered vulnerabilities of PyTorch.
+Beware that none of the topics under [Using Pytorch Securely](#using-pytorch-securely) are considered vulnerabilities of Pytorch.
 
 However, if you believe you have found a security vulnerability in PyTorch, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.
 
 Please report security issues using https://github.com/pytorch/pytorch/security/advisories/new
 
-All reports submitted through the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
+All reports submitted thru the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
 
 Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:
 
 https://www.facebook.com/whitehat
 
 
-## Using PyTorch Securely
-**PyTorch models are programs**, so treat its security seriously -- running untrusted models is equivalent to running untrusted code. In general we recommend that model weights and the python code for the model are distributed independently. That said, be careful about where you get the python code from and who wrote it (preferentially check for a provenance or checksums, do not run any pip installed package).
+## Using Pytorch Securely
+**Pytorch models are programs**, so treat its security seriously -- running untrusted models is equivalent to running untrusted code. In general we recommend that model weights and the python code for the model are distributed independently. That said, be careful about where you get the python code from and who wrote it (preferentially check for a provenance or checksums, do not run any pip installed package).
 
 ### Untrusted models
 Be careful when running untrusted models. This classification includes models created by unknown developers or utilizing data obtained from unknown sources[^data-poisoning-sources].
 
 **Prefer to execute untrusted models within a secure, isolated environment such as a sandbox** (e.g., containers, virtual machines). This helps protect your system from potentially malicious code. You can find further details and instructions in [this page](https://developers.google.com/code-sandboxing).
 
-**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [Safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.
+**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.
 
 Even for more secure serialization formats, unexpected inputs to the downstream system can cause diverse security threats (e.g. denial of service, out of bound reads/writes) and thus we recommend extensive validation of any untrusted inputs.
 
@@ -43,7 +43,7 @@ Important Note: The trustworthiness of a model is not binary. You must always de
 
 ### TorchScript models
 
-TorchScript models should be treated the same way as locally executable code from an unknown source. Only run TorchScript models if you trust the provider. Please note, that tools for introspecting TorchScript models (such as `torch.utils.model_dump`) may also execute partial or full code stored in those models, therefore they should be used only if you trust the provider of the binary you are about to load.
+TorchScript models should treated the same way as locally executable code from an unknown source. Only run TorchScript models if you trust the provider. Please note, that tools for introspecting TorchScript models (such as `torch.utils.model_dump`) may also execute partial or full code stored in those models, therefore they should be used only if you trust the provider of the binary you are about to load.
 
 ### Untrusted inputs during training and prediction
 
@@ -59,9 +59,9 @@ If applicable, prepare your model against bad inputs and prompt injections. Some
 
 ### Data privacy
 
-**Take special security measures if you train your models with sensitive data**. Prioritize [sandboxing](https://developers.google.com/code-sandboxing) your models and:
-- Do not feed sensitive data to an untrusted model (even if runs in a sandboxed environment)
-- If you consider publishing a model that was partially trained with sensitive data, be aware that data can potentially be recovered from the trained weights (especially if the model overfits).
+**Take special security measures if your model if you train models with sensitive data**. Prioritize [sandboxing](https://developers.google.com/code-sandboxing) your models and:
+- Do not feed sensitive data to untrusted model (even if runs in a sandboxed environment)
+- If you consider publishing a model that was partially trained with sensitive data, be aware that data can potentially be recovered from the trained weights (especially if model overfits).
 
 ### Using distributed features
 

aten/src/ATen/Context.cpp

@@ -23,6 +23,8 @@ C10_DIAGNOSTIC_POP()
 #endif
 namespace at {
 
+namespace {
+
 /*
   These const variables defined the fp32 precisions for different backend
   We have "generic", "cuda", "mkldnn" backend now and we can choose fp32
@@ -39,6 +41,16 @@ namespace at {
                 ->rnn
 */
 
+  C10_ALWAYS_INLINE void warn_deprecated_fp32_precision_api(){
+    TORCH_WARN_ONCE(
+      "Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' "
+      "or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, "
+      "torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see "
+      "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices"
+    );
+  }
+} // namespace
+
 Float32Backend str2backend(const std::string& name) {
   if (name == "generic")
     return Float32Backend::GENERIC;
@@ -194,6 +206,7 @@ bool Context::allowTF32CuDNN(std::optional<Float32Op> op) const {
   } else {
     return float32Precision(Float32Backend::CUDA, op.value()) == Float32Precision::TF32;
   }
+  warn_deprecated_fp32_precision_api();
   return allow_tf32_cudnn;
 }
 
@@ -201,6 +214,7 @@ void Context::setAllowTF32CuDNN(bool b) {
   setFloat32Precision(Float32Backend::CUDA, Float32Op::RNN, b ? Float32Precision::TF32 : Float32Precision::NONE);
   setFloat32Precision(Float32Backend::CUDA, Float32Op::CONV, b ? Float32Precision::TF32 : Float32Precision::NONE);
   allow_tf32_cudnn = b;
+  warn_deprecated_fp32_precision_api();
 }
 
 void Context::setSDPPriorityOrder(const std::vector<int64_t>& order) {
@@ -311,6 +325,7 @@ bool Context::allowTF32CuBLAS() const {
       "Current status indicate that you have used mix of the legacy and new APIs to set the TF32 status for cublas matmul. ",
       "We suggest only using the new API to set the TF32 flag. See also: ",
       "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices");
+  warn_deprecated_fp32_precision_api();
   return allow_tf32_new;
 }
 
@@ -334,6 +349,7 @@ Float32MatmulPrecision Context::float32MatmulPrecision() const {
       "Current status indicate that you have used mix of the legacy and new APIs to set the matmul precision. ",
       "We suggest only using the new API for matmul precision. See also: ",
       "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices");
+  warn_deprecated_fp32_precision_api();
   return float32_matmul_precision;
 }
 
@@ -361,6 +377,7 @@ Float32Precision Context::float32Precision(Float32Backend backend, Float32Op op)
 
 void Context::setFloat32MatmulPrecision(const std::string &s) {
   auto match = [this](const std::string & s_) {
+    warn_deprecated_fp32_precision_api();
     // TODO: consider if CuDNN field needs to also be set for potential future CuDNN ops like multi-headed attention
     if (s_ == "highest") {
       float32_matmul_precision = at::Float32MatmulPrecision::HIGHEST;

docs/source/conf.py

@@ -206,41 +206,6 @@ templates_path = [
     os.path.join(os.path.dirname(pytorch_sphinx_theme2.__file__), "templates"),
 ]
 # TODO: document these and remove them from here.
-# Fixes the duplicated
-autosummary_filename_map = {
-    "torch.nn.utils.prune.identity": "torch.nn.utils.prune.identity_function",
-    "torch.nn.utils.prune.Identity": "torch.nn.utils.prune.Identity_class",
-    "torch.optim.adamw.adamw": "torch.optim.adamw.adamw_function",
-    "torch.optim.adamw.AdamW": "torch.optim.adamw.AdamW_class",
-    "torch.optim.asgd.asgd": "torch.optim.asgd.asgd_function",
-    "torch.optim.asgd.ASGD": "torch.optim.asgd.ASGD_class",
-    "torch.optim.nadam.nadam": "torch.optim.nadam.nadam_function",
-    "torch.optim.nadam.NAdam": "torch.optim.nadam.NAdam_class",
-    "torch.optim.radam.radam": "torch.optim.radam.radam_function",
-    "torch.optim.radam.RAdam": "torch.optim.radam.RAdam_class",
-    "torch.optim.rmsprop.rmsprop": "torch.optim.rmsprop.rmsprop_function",
-    "torch.optim.rmsprop.RMSprop": "torch.optim.rmsprop.RMSprop_class",
-    "torch.optim.rprop.rprop": "torch.optim.rprop.rprop_function",
-    "torch.optim.rprop.Rprop": "torch.optim.rprop.Rprop_class",
-    "torch.optim.sgd.sgd": "torch.optim.sgd.sgd_function",
-    "torch.optim.sgd.SGD": "torch.optim.sgd.SGD_class",
-    "torch.optim.adadelta.adadelta": "torch.optim.adadelta.adadelta_function",
-    "torch.optim.adadelta.Adadelta": "torch.optim.adadelta.Adadelta_class",
-    "torch.optim.adagrad.adagrad": "torch.optim.adagrad.adagrad_function",
-    "torch.optim.adagrad.Adagrad": "torch.optim.adagrad.Adagrad_class",
-    "torch.optim.adam.adam": "torch.optim.adam.adam_function",
-    "torch.optim.adam.Adam": "torch.optim.adam.Adam_class",
-    "torch.optim.adamax.adamax": "torch.optim.adamax.adamax_function",
-    "torch.optim.adamax.Adamax": "torch.optim.adamax.Adamax_class",
-    "torch.mtia.stream": "torch.mtia.stream_function",
-    "torch.mtia.Stream": "torch.mtia.Stream_class",
-    "torch.cpu.stream": "torch.cpu.stream_function",
-    "torch.cpu.Stream": "torch.cpu.Stream_class",
-    "torch.cuda.stream": "torch.cuda.stream_function",
-    "torch.cuda.Stream": "torch.cuda.Stream_class",
-    "torch.xpu.stream": "torch.xpu.stream_function",
-    "torch.xpu.Stream": "torch.xpu.Stream_class",
-}
 
 coverage_ignore_functions = [
     # torch
@@ -3230,11 +3195,6 @@ autodoc_type_aliases = {
 # Enable overriding of function signatures in the first line of the docstring.
 autodoc_docstring_signature = True
 
-# Exclude inherited IntEnum methods that have RST formatting issues in their docstrings
-autodoc_default_options = {
-    "exclude-members": "from_bytes, to_bytes",
-}
-
 # -- katex javascript in header
 #
 #    def setup(app):

docs/source/quantization-support.md

@@ -253,6 +253,7 @@ regular full-precision tensor.
 .. autosummary::
     :toctree: generated
     :nosignatures:
+    :template: classtemplate.rst
 
     view
     as_strided

test/inductor/test_compile_worker.py

@@ -4,7 +4,6 @@ import os
 import tempfile
 from threading import Event
 
-import torch._inductor.config as config
 from torch._inductor.compile_worker.subproc_pool import (
     raise_testexc,
     SubprocException,
@@ -17,12 +16,9 @@ from torch.testing._internal.inductor_utils import HAS_CPU
 
 
 class TestCompileWorker(TestCase):
-    def make_pool(self, size):
-        return SubprocPool(size)
-
     @skipIfWindows(msg="pass_fds not supported on Windows.")
     def test_basic_jobs(self):
-        pool = self.make_pool(2)
+        pool = SubprocPool(2)
         try:
             a = pool.submit(operator.add, 100, 1)
             b = pool.submit(operator.sub, 100, 1)
@@ -33,7 +29,7 @@ class TestCompileWorker(TestCase):
 
     @skipIfWindows(msg="pass_fds not supported on Windows.")
     def test_exception(self):
-        pool = self.make_pool(2)
+        pool = SubprocPool(2)
         try:
             a = pool.submit(raise_testexc)
             with self.assertRaisesRegex(
@@ -46,7 +42,7 @@ class TestCompileWorker(TestCase):
 
     @skipIfWindows(msg="pass_fds not supported on Windows.")
     def test_crash(self):
-        pool = self.make_pool(2)
+        pool = SubprocPool(2)
         try:
             with self.assertRaises(Exception):
                 a = pool.submit(os._exit, 1)
@@ -62,7 +58,7 @@ class TestCompileWorker(TestCase):
 
     @skipIfWindows(msg="pass_fds not supported on Windows.")
     def test_quiesce(self):
-        pool = self.make_pool(2)
+        pool = SubprocPool(2)
         try:
             a = pool.submit(operator.add, 100, 1)
             pool.quiesce()
@@ -79,7 +75,7 @@ class TestCompileWorker(TestCase):
         os.environ["ROLE_RANK"] = "0"
         with tempfile.NamedTemporaryFile(delete=True) as temp_log:
             os.environ["TORCHINDUCTOR_WORKER_LOGPATH"] = temp_log.name
-            pool = self.make_pool(2)
+            pool = SubprocPool(2)
             try:
                 pool.submit(operator.add, 100, 1)
                 self.assertEqual(os.path.exists(temp_log.name), True)
@@ -87,12 +83,6 @@ class TestCompileWorker(TestCase):
                 pool.shutdown()
 
 
-@config.patch("quiesce_async_compile_time", 0.1)
-class TestCompileWorkerWithTimer(TestCompileWorker):
-    def make_pool(self, size):
-        return SubprocPool(size, quiesce=True)
-
-
 class TestTimer(TestCase):
     def test_basics(self):
         done = Event()

test/inductor/test_padding.py

@@ -500,13 +500,8 @@ class PaddingTest(TestCaseBase):
         forward_wrapper = wrapper_codes[0]
 
         # make sure the load for softmax is aligned
-        if bias:
-            # addmm -> mm + bias and bias is fused with softmax
-            softmax_load_str = "tl.load(in_out_ptr0 + (r0_1 + 30528*x0)"
-        else:
-            softmax_load_str = "tl.load(in_ptr0 + (r0_1 + 30528*x0)"
         self.assertTrue(
-            softmax_load_str in forward_wrapper,
+            "tl.load(in_ptr0 + (r0_1 + 30528*x0)" in forward_wrapper,
             f"forward_wrapper: {forward_wrapper}",
         )
 

test/inductor/test_torchinductor.py

@@ -15280,7 +15280,7 @@ if RUN_GPU:
                 ),
                 (
                     fn3,
-                    "triton_poi_fused_addmm_native_layer_norm",
+                    "triton_poi_fused_native_layer_norm_relu",
                     (torch.randn(4, 4, device=GPU_TYPE),),
                 ),
             ]
@@ -15293,7 +15293,7 @@ if RUN_GPU:
                 ),
                 (
                     fn3,
-                    "triton_poi_fused_LayerNorm_Linear_ReLU",
+                    "triton_poi_fused_LayerNorm_ReLU",
                     (torch.randn(4, 4, device=GPU_TYPE),),
                 ),
             ]

test/run_test.py

@@ -1826,14 +1826,9 @@ def run_test_module(
         test_name = test.name
 
         # Printing the date here can help diagnose which tests are slow
-        start = time.perf_counter()
-        print_to_stderr(f"Running {str(test)} ... [{datetime.now()}][{start}]")
+        print_to_stderr(f"Running {str(test)} ... [{datetime.now()}]")
         handler = CUSTOM_HANDLERS.get(test_name, run_test)
         return_code = handler(test, test_directory, options)
-        end = time.perf_counter()
-        print_to_stderr(
-            f"Finished {str(test)} ... [{datetime.now()}][{end}], took {(end - start) / 60:.2f}min"
-        )
         assert isinstance(return_code, int) and not isinstance(return_code, bool), (
             f"While running {str(test)} got non integer return code {return_code}"
         )

test/test_cuda.py

@@ -7413,140 +7413,6 @@ class TestCudaDeviceParametrized(TestCase):
         )
 
 
-class TestFXMemoryProfiler(TestCase):
-    """Tests for memory profiler augmentation with original stack traces."""
-
-    def collect_frames(
-        self, augmented_snapshot, collect_device_traces=True, collect_segments=True
-    ):
-        """Collects all frames that has node metadata from a memory snapshot."""
-        # Collect all frames with FX metadata
-        fx_frames = []
-
-        # Check device traces for FX debug fields
-        if collect_device_traces and "device_traces" in augmented_snapshot:
-            for trace_list in augmented_snapshot["device_traces"]:
-                for trace_entry in trace_list:
-                    if isinstance(trace_entry, dict) and "frames" in trace_entry:
-                        for frame in trace_entry["frames"]:
-                            if isinstance(frame, dict):
-                                # Check for FX debug fields
-                                if "fx_node_op" in frame or "fx_node_name" in frame:
-                                    fx_frames.append(frame)
-
-        # Check segments/blocks for FX debug fields
-        if collect_segments and "segments" in augmented_snapshot:
-            for segment in augmented_snapshot["segments"]:
-                if "blocks" in segment:
-                    for block in segment["blocks"]:
-                        if "frames" in block:
-                            for frame in block["frames"]:
-                                if isinstance(frame, dict):
-                                    if "fx_node_op" in frame or "fx_node_name" in frame:
-                                        fx_frames.append(frame)
-        return fx_frames
-
-    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available")
-    @torch._dynamo.config.patch("enrich_profiler_metadata", True)
-    def test_fx_memory_profiler_augmentation(self):
-        """Test that memory snapshots are augmented with FX debug information."""
-
-        # Create a simple model
-        class MLPModule(nn.Module):
-            def __init__(self, device):
-                super().__init__()
-                torch.manual_seed(5)
-                self.net1 = nn.Linear(10, 16, bias=True, device=device)
-                self.relu = nn.ReLU()
-                self.net2 = nn.Linear(16, 10, bias=True, device=device)
-
-            def forward(self, x):
-                a = self.net1(x)
-                b = self.relu(a)
-                c = self.net2(b)
-                return c
-
-        device = "cuda"
-        mod = MLPModule(device)
-        with tempfile.TemporaryDirectory() as tmpdir:
-            torch.cuda.memory._record_memory_history()
-            compiled = torch.compile(mod, backend="aot_eager", fullgraph=True)
-            result = compiled(torch.randn(10, 10, device=device))
-            augmented_snapshot = torch.cuda.memory._snapshot(
-                augment_with_fx_traces=True
-            )
-            torch.cuda.memory._record_memory_history(enabled=None, clear_history=True)
-            torch.cuda.empty_cache()
-
-            fx_frames = self.collect_frames(augmented_snapshot)
-            if TEST_WITH_ROCM:
-                self.assertGreater(len(fx_frames), 0)
-            else:
-                self.assertEqual(len(fx_frames), 12)
-
-            for frame in fx_frames:
-                # Every FX frame should have both node_op and node_name
-                self.assertIn("fx_node_op", frame)
-                self.assertIn("fx_node_name", frame)
-                self.assertIn("fx_node_target", frame)
-                self.assertIn("fx_original_trace", frame)
-
-                self.assertIn(frame["fx_node_name"], ["addmm", "relu", "addmm_1"])
-                fx_node_name = frame["fx_node_name"]
-                if fx_node_name == "addmm":
-                    self.assertIn("a = self.net1(x)", frame["fx_original_trace"])
-                elif fx_node_name == "addmm_1":
-                    self.assertIn("c = self.net2(b)", frame["fx_original_trace"])
-                elif fx_node_name == "relu":
-                    self.assertIn("b = self.relu(a)", frame["fx_original_trace"])
-
-        # Test that when we have two graphs with the same src_code, they're not hashed
-        # to the same metadata
-        class MLPModule2(nn.Module):
-            def __init__(self, device):
-                super().__init__()
-                torch.manual_seed(5)
-                self.net1 = nn.Linear(10, 16, bias=True, device=device)
-                self.relu = nn.ReLU()
-                self.net2 = nn.Linear(16, 10, bias=True, device=device)
-
-            def forward(self, x):
-                d = self.net1(x)
-                e = self.relu(d)
-                f = self.net2(e)
-                return f
-
-        mod = MLPModule2(device)
-        with tempfile.TemporaryDirectory() as tmpdir:
-            torch.cuda.memory._record_memory_history()
-            compiled = torch.compile(mod, backend="aot_eager", fullgraph=True)
-            result = compiled(torch.randn(10, 10, device=device))
-            augmented_snapshot = torch.cuda.memory._snapshot(
-                augment_with_fx_traces=True
-            )
-            torch.cuda.memory._record_memory_history(enabled=None, clear_history=True)
-
-            # avoid collecting segments from previous run for unit test purpose
-            fx_frames = self.collect_frames(augmented_snapshot, collect_segments=False)
-            self.assertGreater(len(fx_frames), 0)
-
-            for frame in fx_frames:
-                # Every FX frame should have both node_op and node_name
-                self.assertIn("fx_node_op", frame)
-                self.assertIn("fx_node_name", frame)
-                self.assertIn("fx_node_target", frame)
-                self.assertIn("fx_original_trace", frame)
-
-                self.assertIn(frame["fx_node_name"], ["addmm", "relu", "addmm_1"])
-                fx_node_name = frame["fx_node_name"]
-                if fx_node_name == "addmm":
-                    self.assertIn("d = self.net1(x)", frame["fx_original_trace"])
-                elif fx_node_name == "addmm_1":
-                    self.assertIn("f = self.net2(e)", frame["fx_original_trace"])
-                elif fx_node_name == "relu":
-                    self.assertIn("e = self.relu(d)", frame["fx_original_trace"])
-
-
 instantiate_parametrized_tests(TestCuda)
 instantiate_parametrized_tests(TestCudaMallocAsync)
 instantiate_parametrized_tests(TestCompileKernel)

test/test_fx.py

@@ -771,7 +771,6 @@ class TestFX(JitTestCase):
         gm = GraphModule(tracer.root, graph)
         expected = {1: 2, 2: 3, 3: 4, 4: 5}
         self.assertTrue(set(expected.items()).issubset(set(gm._lineno_map.items())))
-        self.assertEqual(gm._prologue_start, 4)
 
         # test custom codegen
         def transform_code(code):
@@ -781,7 +780,6 @@ class TestFX(JitTestCase):
         gm.recompile()
         expected = {2: 2, 3: 3, 4: 4, 5: 5}
         self.assertTrue(set(expected.items()).issubset(set(gm._lineno_map.items())))
-        self.assertEqual(gm._prologue_start, 4)
 
     def test_graph_unique_names_manual(self):
         graph: torch.fx.Graph = torch.fx.Graph()

torch/_dynamo/config.py

@@ -739,12 +739,6 @@ enable_aot_compile = False
 # HACK: this is for testing custom ops profiling only
 _custom_ops_profile: Optional[Any] = None
 
-# Experimental: If True, graph module will register fx metadata during recompile()
-enrich_profiler_metadata: bool = Config(  # type: ignore[var-annotated]
-    default=False,
-    env_name_default="TORCH_ENRICH_RPOFILER_STACK_TRACE",
-)
-
 if TYPE_CHECKING:
     from torch.utils._config_typing import *  # noqa: F401, F403
 

torch/_inductor/compile_worker/subproc_pool.py

@@ -24,7 +24,6 @@ from typing_extensions import Never, ParamSpec
 import torch._thread_safe_fork  # noqa: F401
 from torch._inductor import config
 from torch._inductor.codecache import torch_key
-from torch._inductor.compile_worker.timer import Timer
 from torch._inductor.compile_worker.tracked_process_pool import (
     TrackedProcessPoolExecutor,
 )
@@ -133,7 +132,6 @@ class SubprocPool:
         nprocs: int,
         pickler: Optional[SubprocPickler] = None,
         kind: SubprocKind = SubprocKind.FORK,
-        quiesce: bool = False,
     ) -> None:
         entry = os.path.join(os.path.dirname(__file__), "__main__.py")
         self.pickler = pickler or SubprocPickler()
@@ -218,13 +216,6 @@ class SubprocPool:
             "pytorch.wait_counter.subproc_pool.first_job"
         ).guard()
 
-        if quiesce:
-            self.timer: Optional[Timer] = Timer(
-                config.quiesce_async_compile_time, self.quiesce
-            )
-        else:
-            self.timer = None
-
         # Start thread last to ensure all member variables are initialized
         # before any access.
         self.read_thread.start()
@@ -297,8 +288,6 @@ class SubprocPool:
             with self.futures_lock:
                 if not self.running:
                     return
-                if self.timer:
-                    self.timer.record_call()
                 if isinstance(result, _SubprocExceptionInfo):
                     # An exception occurred in the submitted job
                     self.pending_futures[job_id].set_exception(
@@ -333,8 +322,6 @@ class SubprocPool:
             with self.write_lock:
                 if not self.running:
                     return
-                if self.timer:
-                    self.timer.quit()
                 self.running = False
                 self.running_waitcounter.__exit__()
                 _send_msg(self.write_pipe, MsgHeader.SHUTDOWN)

torch/_inductor/compile_worker/timer.py

@@ -17,7 +17,7 @@ class Timer:
         self.background_thread: Optional[Thread] = None
         self.last_called: Optional[float] = None
         self.duration = duration
-        self.sleep_time = duration / 2
+        self.sleep_time = 60
         self.call = call
         self.exit = False
 

torch/_inductor/config.py

@@ -964,11 +964,6 @@ quiesce_async_compile_pool: bool = Config(
     default=False,
 )
 
-# Time in seconds to wait before quiescing
-quiesce_async_compile_time: int = Config(
-    default=60,
-)
-
 # Whether or not to enable statically launching CUDA kernels
 # compiled by triton (instead of using triton's own launcher)
 use_static_cuda_launcher: bool = static_cuda_launcher_default()

torch/_inductor/fx_passes/post_grad.py

@@ -51,8 +51,8 @@ from ..utils import (
     decode_device,
     get_all_devices,
     get_gpu_type,
-    has_uses_tagged_as,
     is_gpu,
+    is_pointwise_use,
     OPTIMUS_EXCLUDE_POST_GRAD,
 )
 from ..virtualized import V
@@ -1510,10 +1510,8 @@ def should_prefer_unfused_addmm(match):
     if not is_gpu(inp.meta["val"].device.type):
         return False
 
-    return has_uses_tagged_as(
-        match.output_node(),
-        (torch.Tag.pointwise, torch.Tag.reduction),
-    )
+    output = match.output_node()
+    return all(is_pointwise_use(use) for use in output.users)
 
 
 @register_graph_pattern(

torch/_inductor/utils.py

@@ -549,70 +549,6 @@ def is_pointwise_use(
     return torch.Tag.pointwise in target.tags or is_pointwise_fn(target)
 
 
-class LogicalConnective(enum.Enum):
-    OR = enum.auto()
-    AND = enum.auto()
-
-
-def has_uses(
-    target: Node,
-    use_selector_fn: Callable[[torch._ops.OpOverload], bool] = lambda _: False,
-    use_aggregate_type: LogicalConnective = LogicalConnective.OR,
-) -> bool:
-    """
-    Given a target, explore the uses of `target` by applying `use_selector_fn`
-    on them, and then aggregate these booleans with the `use_aggregate_type`
-    logical connective.
-
-    Uses in view ops will follow the views uses.
-    """
-
-    def get_use_aggregate_fn(
-        use_aggregate_type: LogicalConnective,
-    ) -> Callable[[Iterator[Any]], bool]:
-        match use_aggregate_type:
-            case LogicalConnective.AND:
-                return all
-            case LogicalConnective.OR:
-                return any
-            case _:
-                return any
-
-    use_aggregate_fn = get_use_aggregate_fn(use_aggregate_type)
-
-    def has_uses_impl(use: Node) -> bool:
-        if use.op != "call_function":
-            return False
-        if not (
-            isinstance(use.target, torch._ops.OpOverload)
-            or use.target is operator.getitem
-        ):
-            return False
-
-        target = cast(torch._ops.OpOverload, use.target)
-        # Process getitem and view
-        if target is operator.getitem or is_view(target):
-            return use_aggregate_fn(has_uses_impl(user) for user in use.users)
-
-        return use_selector_fn(target)
-
-    return use_aggregate_fn(has_uses_impl(user) for user in target.users)
-
-
-def has_uses_tagged_as(
-    target: Node,
-    use_tags: Collection[torch.Tag],
-    use_aggregate_type: LogicalConnective = LogicalConnective.OR,
-) -> bool:
-    """
-    Is there a use with given tags?
-    """
-
-    return has_uses(
-        target, lambda use: any(tag in use_tags for tag in use.tags), use_aggregate_type
-    )
-
-
 def gen_gm_and_inputs(
     target: Any, args: list[Any], kwargs: dict[str, Any]
 ) -> tuple[GraphModule, list[torch.Tensor]]:
@@ -2001,10 +1937,11 @@ def use_blackwell_cutedsl_grouped_mm(
     """
     Returns True if we can use the blackwell kernel for grouped mm.
     Required conditions:
-        1. CuTeDSL is available
-        2. We are on a blackwell arch
-        3. The dtype is bf16
-        4. Max autotune or max autotune gemm is enabled
+        1. CuTeDSL backend is enabled
+        2. CuTeDSL is available
+        3. We are on a blackwell arch
+        4. The dtype is bf16
+        5. Max autotune or max autotune gemm is enabled
         6. A, B, and the output are 16B aligned
         7. We are not using dynamic shapes
         8. A is 2d
@@ -2015,9 +1952,15 @@ def use_blackwell_cutedsl_grouped_mm(
     if not ensure_cute_available():
         return False
 
+    if not _use_autotune_backend("CUTEDSL"):
+        return False
+
     from .codegen.cuda.cuda_env import is_datacenter_blackwell_arch
 
-    if not is_gpu(layout.device.type) and is_datacenter_blackwell_arch():
+    if not is_gpu(layout.device.type):
+        return False
+
+    if not is_datacenter_blackwell_arch():
         return False
 
     layout_dtypes = [torch.bfloat16]

torch/csrc/stable/stableivalue_conversions.h

@@ -31,8 +31,10 @@ template <typename T>
 struct FromImpl {
   static StableIValue call(
       T val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     static_assert(
         sizeof(T) <= sizeof(StableIValue),
         "StableLibrary stack does not support parameter types larger than 64 bits.");
@@ -73,8 +75,10 @@ template <>
 struct FromImpl<ScalarType> {
   static StableIValue call(
       ScalarType val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     switch (val) {
       case ScalarType::Byte:
         return from(aoti_torch_dtype_uint8());
@@ -129,8 +133,10 @@ template <>
 struct FromImpl<std::nullopt_t> {
   static StableIValue call(
       std::nullopt_t val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     return from(nullptr);
   }
 };
@@ -184,8 +190,10 @@ template <>
 struct FromImpl<torch::stable::Tensor> {
   static StableIValue call(
       const torch::stable::Tensor& val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     AtenTensorHandle new_ath;
     TORCH_ERROR_CODE_CHECK(aoti_torch_new_tensor_handle(val.get(), &new_ath));
     return from(new_ath);
@@ -201,8 +209,10 @@ template <typename T>
 struct ToImpl {
   static T call(
       StableIValue val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     static_assert(std::is_trivially_copyable_v<T>);
     // T may not have a default constructor. (For example, it might be
     // c10::Device.) However, std::memcpy implicitly creates a T at the
@@ -239,8 +249,10 @@ template <>
 struct ToImpl<ScalarType> {
   static ScalarType call(
       StableIValue val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     int32_t shim_scalartype = to<int32_t>(val);
     if (shim_scalartype == aoti_torch_dtype_uint8()) {
       return ScalarType::Byte;
@@ -297,8 +309,10 @@ template <>
 struct ToImpl<std::nullopt_t> {
   static std::nullopt_t call(
       StableIValue val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     // val should be equivalent to from(nullptr)
     return std::nullopt;
   }
@@ -336,8 +350,10 @@ template <>
 struct ToImpl<torch::stable::Tensor> {
   static torch::stable::Tensor call(
       StableIValue val,
-      [[maybe_unused]] uint64_t extension_build_version,
-      [[maybe_unused]] bool is_internal) {
+      uint64_t extension_build_version,
+      bool is_internal) {
+    (void)extension_build_version; // Unused parameter
+    (void)is_internal; // Unused parameter
     return torch::stable::Tensor(to<AtenTensorHandle>(val));
   }
 };

torch/cuda/memory.py

@@ -4,14 +4,12 @@ r"""This package adds support for device memory management implemented in CUDA."
 import collections
 import contextlib
 import ctypes
-import os
 import pickle
-import re
 import sys
 import warnings
 from inspect import signature
-from typing import Any, cast, Literal, Optional, TYPE_CHECKING, TypedDict
-from typing_extensions import deprecated, NotRequired
+from typing import Any, Literal, Optional, TYPE_CHECKING
+from typing_extensions import deprecated
 
 import torch
 from torch import _C
@@ -31,60 +29,6 @@ if TYPE_CHECKING:
     from torch.types import Device
 
 
-# Type definitions for memory profiler
-class _Frame(TypedDict):
-    """Frame information from memory profiler snapshots."""
-
-    filename: str
-    line: int
-    name: str
-    # Fields added by FX augmentation (optional)
-    fx_node_op: NotRequired[str]
-    fx_node_name: NotRequired[str]
-    fx_node_target: NotRequired[str]
-    fx_original_trace: NotRequired[str]
-
-
-class _Block(TypedDict):
-    """Memory block information."""
-
-    size: int
-    requested_size: int
-    address: int
-    state: str
-    frames: list[_Frame]
-
-
-class _Segment(TypedDict):
-    """Memory segment information."""
-
-    address: int
-    total_size: int
-    stream: int
-    segment_type: str
-    allocated_size: int
-    active_size: int
-    blocks: list[_Block]
-
-
-class _TraceEntry(TypedDict):
-    """Memory trace entry information."""
-
-    action: str
-    addr: NotRequired[int]
-    frames: list[_Frame]
-    size: int
-    stream: int
-    device_free: NotRequired[int]
-
-
-class _Snapshot(TypedDict):
-    """Memory snapshot structure."""
-
-    segments: list[_Segment]
-    device_traces: NotRequired[list[list[_TraceEntry]]]
-
-
 __all__ = [
     "caching_allocator_alloc",
     "caching_allocator_delete",
@@ -1020,120 +964,7 @@ def _record_memory_history_impl(
 _record_memory_history.__signature__ = signature(_record_memory_history_impl)  # type: ignore[attr-defined]
 
 
-def _augment_frames(frames: list[_Frame]) -> int:
-    """
-    Augment a list of frames with FX debug information.
-
-    Args:
-        frames: List of frame dictionaries to augment
-
-    Returns:
-        The count of frames that were augmented.
-    """
-    from torch.fx.graph_module import FX_GRAPH_MODULE_FILE_PREFIX
-
-    # Regex pattern to match FX generated files
-    _FX_GENERATED_PATTERN = re.compile(
-        rf"{re.escape(FX_GRAPH_MODULE_FILE_PREFIX)}.*\.py$"
-    )
-
-    count = 0
-    if not frames:
-        return count
-
-    for frame in frames:
-        if "filename" in frame and "line" in frame:
-            filename = frame["filename"]
-            lineno = frame["line"]
-
-            # Check if this looks like an FX generated file
-            if not _FX_GENERATED_PATTERN.search(os.path.basename(filename)):
-                continue
-
-            # Look up metadata from the global registry
-            from torch.fx.traceback import _FX_METADATA_REGISTRY
-
-            metadata = _FX_METADATA_REGISTRY.get(filename)
-            if metadata is None:
-                continue
-
-            lineno_map = metadata.get("lineno_map", {})
-            node_metadata = metadata.get("node_metadata", {})
-            prologue_start = metadata.get("prologue_start", 0)
-
-            # Get the node index for this line
-            node_idx = lineno_map.get(lineno - prologue_start)
-
-            if node_idx is not None and node_idx in node_metadata:
-                node_info = node_metadata[node_idx]
-                original_trace = node_info.get("stack_trace")
-                node_op = node_info.get("op")
-                node_name = node_info.get("name")
-                node_target = node_info.get("target")
-
-                # Always add node metadata
-                frame["fx_node_op"] = node_op
-                frame["fx_node_name"] = node_name
-                frame["fx_node_target"] = str(node_target)
-
-                # Add original trace if available
-                if original_trace:
-                    frame["fx_original_trace"] = original_trace
-
-                count += 1
-
-    return count
-
-
-def _augment_memory_snapshot_stack_traces(
-    snapshot: str | _Snapshot,
-) -> _Snapshot:
-    """
-    Augment a memory snapshot with original source stack traces from FX metadata.
-
-    IMPORTANT: This function reads from a global in-memory registry (_FX_METADATA_REGISTRY)
-    that is populated during graph module compilation. It must be called in the same
-    Python process where the FX graphs were compiled. It cannot be used to augment
-    snapshots loaded from disk in a different process.
-
-    Args:
-        snapshot: Either a memory snapshot dict or path to a snapshot pickle file
-
-    Returns:
-        The augmented snapshot dictionary with fx_node_op, fx_node_name,
-        fx_original_trace, and fx_node_info fields added to frames
-    """
-
-    snapshot_dict: _Snapshot
-    if isinstance(snapshot, str):
-        # Load the memory snapshot
-        with open(snapshot, "rb") as f:
-            snapshot_dict = cast(_Snapshot, pickle.load(f))
-    else:
-        snapshot_dict = snapshot
-
-    # Process stack traces in the snapshot
-    augmented_count = 0
-
-    # Process blocks in segments (for regular allocations)
-    if "segments" in snapshot_dict:
-        for segment in snapshot_dict["segments"]:
-            if "blocks" in segment:
-                for block in segment["blocks"]:
-                    if "frames" in block:
-                        augmented_count += _augment_frames(block["frames"])
-
-    # Process device traces (for memory history)
-    if "device_traces" in snapshot_dict:
-        for trace_list in snapshot_dict["device_traces"]:
-            for trace_entry in trace_list:
-                if isinstance(trace_entry, dict) and "frames" in trace_entry:
-                    augmented_count += _augment_frames(trace_entry["frames"])
-
-    return snapshot_dict
-
-
-def _snapshot(device: "Device" = None, augment_with_fx_traces=False):
+def _snapshot(device: "Device" = None):
     """Save a snapshot of CUDA memory state at the time it was called.
 
     The state is represented as a dictionary with the following structure.
@@ -1181,11 +1012,6 @@ def _snapshot(device: "Device" = None, augment_with_fx_traces=False):
             filename: str
             line: int
             name: str
-            # Optional FX debug fields (present when augment_with_fx_traces=True
-            # and the frame corresponds to FX-generated code)
-            fx_node_op: str  # FX node operation type (e.g., 'call_function', 'output')
-            fx_node_name: str  # FX node name (e.g., 'linear', 'relu_1')
-            fx_original_trace: str  # Original model source code stack trace
 
 
         class TraceEntry(TypedDict):
@@ -1215,23 +1041,13 @@ def _snapshot(device: "Device" = None, augment_with_fx_traces=False):
             device_free: int  # only present for OOM, the amount of
             # memory cuda still reports to be free
 
-    Args:
-        device: Device to capture snapshot for. If None, captures for current device.
-        augment_with_fx_traces: If True, augment stack trace frames with FX debug information
-                                that maps generated FX code back to original model source code.
-                                This adds fx_node_op, fx_node_name, fx_original_trace, and
-                                fx_node_info fields to Frame objects. Default: False.
-
     Returns:
         The Snapshot dictionary object
     """
-    s = _C._cuda_memorySnapshot(None)
-    if augment_with_fx_traces:
-        s = _augment_memory_snapshot_stack_traces(s)  # type: ignore[assignment, arg-type]
-    return s
+    return _C._cuda_memorySnapshot(None)
 
 
-def _dump_snapshot(filename="dump_snapshot.pickle", augment_with_fx_traces=False):
+def _dump_snapshot(filename="dump_snapshot.pickle"):
     """
     Save a pickled version of the `torch.memory._snapshot()` dictionary to a file.
 
@@ -1243,14 +1059,8 @@ def _dump_snapshot(filename="dump_snapshot.pickle", augment_with_fx_traces=False
 
     Args:
         filename (str, optional): Name of the file to create. Defaults to "dump_snapshot.pickle".
-        augment_with_fx_traces (bool, optional): If True, augment the snapshot with FX debug information
-                                                  before dumping. This maps generated FX code stack traces
-                                                  back to original model source code. Defaults to False.
-        verbose (bool, optional): If True and augment_with_fx_traces is True, print verbose debug output
-                                  during augmentation. Defaults to False.
     """
-    s = _snapshot(augment_with_fx_traces=augment_with_fx_traces)
-
+    s = _snapshot()
     with open(filename, "wb") as f:
         pickle.dump(s, f)
 

torch/fx/graph.py

@@ -226,10 +226,8 @@ class PythonCode:
     # Values in global scope during execution of `src_def`.
     globals: dict[str, Any]
     # Optional mapping from the forward function's line number to
-    # node index. Line number starts at the prologue (i.e. forward()).
+    # node index.
     _lineno_map: Optional[dict[int, Optional[int]]]
-    # The line number of prologue in fn_code
-    _prologue_start: int = 0
 
 
 def _format_target(base: str, target: str) -> str:
@@ -856,14 +854,7 @@ class CodeGen:
 
 {prologue}
 {code}"""
-        # The +4 accounts for the empty lines before prologue in fn_code
-        prologue_start = wrap_stmts.count("\n") + 4
-        return PythonCode(
-            fn_code,
-            globals_,
-            _lineno_map=lineno_map,
-            _prologue_start=prologue_start,
-        )
+        return PythonCode(fn_code, globals_, _lineno_map=lineno_map)
 
 
 # Ideally, we'd like to refactor all of the pytree logic into this codegen

torch/fx/graph_module.py

@@ -1,8 +1,6 @@
 # mypy: allow-untyped-defs
-import base64
 import contextlib
 import copy
-import hashlib
 import itertools
 import linecache
 import os
@@ -38,7 +36,6 @@ __all__ = [
 ]
 
 _USER_PRESERVED_ATTRIBUTES_KEY = "_user_preserved_attributes"
-FX_GRAPH_MODULE_FILE_PREFIX = "fx_generated_"
 
 
 # Normal exec loses the source code, however we can work with
@@ -64,13 +61,7 @@ class _EvalCacheLoader:
 
         key = self._get_key()
         if co_fields:
-            if "co_filename" in co_fields:
-                # If only co_filename is provided, use it directly as the key
-                if "co_firstlineno" not in co_fields or "co_name" not in co_fields:
-                    key = co_fields["co_filename"]
-                else:
-                    # Full co_fields with all three components
-                    key += f" from {co_fields['co_filename']}:{co_fields['co_firstlineno']} in {co_fields['co_name']}"
+            key += f" from {co_fields['co_filename']}:{co_fields['co_firstlineno']} in {co_fields['co_name']}"
         self.eval_cache[key] = src
 
         # Don't mutate globals so that this loader is only used
@@ -362,36 +353,6 @@ def _print_readable(
     return output
 
 
-def _metadata_hash(code: str, node_metadata: dict) -> str:
-    """
-    Create a content-addressed hash from code and metadata.
-
-    Args:
-        code: The source code string
-        lineno_map: Mapping from line numbers to node indices
-        node_metadata: Metadata for each node
-
-    Returns:
-        A 51-character base32-encoded hash
-    """
-    import json
-
-    # Create a deterministic string representation of all components
-    # We use JSON to ensure consistent serialization
-    hash_data = {
-        "code": code,
-        "node_metadata": node_metadata,
-    }
-    hashing_str = json.dumps(hash_data).encode("utf-8")
-
-    # [:51] to strip off the "Q====" suffix common to every hash value.
-    return (
-        base64.b32encode(hashlib.sha256(hashing_str).digest())[:51]
-        .decode("utf-8")
-        .lower()
-    )
-
-
 class _WrappedCall:
     def __init__(self, cls, cls_call):
         self.cls = cls
@@ -864,47 +825,9 @@ class {module_name}(torch.nn.Module):
         python_code = self._graph.python_code(root_module="self")
         self._code = python_code.src
         self._lineno_map = python_code._lineno_map
-        self._prologue_start = python_code._prologue_start
 
         cls = type(self)
         co_fields = self._graph._co_fields if hasattr(self._graph, "_co_fields") else {}
-        from torch._dynamo import config as dynamo_config
-
-        if dynamo_config.enrich_profiler_metadata:
-            # Generate metadata and register for profiler augmentation
-            node_metadata: dict[int, dict[str, Any]] = {}
-            for i, node in enumerate(self._graph.nodes):
-                node_metadata[i] = {
-                    "name": node.name,
-                    "op": node.op,
-                    "target": str(node.target),
-                    "stack_trace": node.meta.get("stack_trace", None),
-                }
-
-            # Generate a content-addressed filename based on hash of code and metadata
-            # This ensures the same code+metadata always generates the same filename
-            hash_value = _metadata_hash(self._code, node_metadata)
-            file_stem = f"{FX_GRAPH_MODULE_FILE_PREFIX}_{hash_value}"
-
-            filename = f"{file_stem}.py"
-
-            # Only include co_filename to use it directly as the cache key
-            co_fields = {
-                "co_filename": filename,
-            }
-
-            # Store metadata in global in-memory registry
-            metadata = {
-                "lineno_map": python_code._lineno_map,
-                "prologue_start": python_code._prologue_start,
-                "node_metadata": node_metadata,
-            }
-
-            # Register metadata in the global registry
-            from torch.fx.traceback import _register_fx_metadata
-
-            _register_fx_metadata(filename, metadata)
-
         cls.forward = _forward_from_src(self._code, python_code.globals, co_fields)
 
         # Determine whether this class explicitly defines a __call__ implementation

torch/fx/traceback.py

@@ -38,28 +38,6 @@ current_meta: dict[str, Any] = {}
 current_replay_node: Optional[Node] = None
 should_preserve_node_meta = False
 
-# =============================================================================
-# FX Metadata Registry for Memory Profiler
-# =============================================================================
-# Global in-memory registry for FX metadata
-# Maps module_name -> metadata dict containing lineno_map and node_metadata
-_FX_METADATA_REGISTRY: dict[str, dict[str, Any]] = {}
-
-
-def _register_fx_metadata(module_name: str, metadata: dict[str, Any]) -> None:
-    """
-    Register FX metadata in the global in-memory registry.
-
-    This is called automatically during graph module compilation to store metadata
-    for later use by memory profiler augmentation.
-
-    Args:
-        module_name: The module identifier (content-addressed filename)
-        metadata: Metadata dict containing lineno_map, node_metadata, and source_code
-    """
-    # TODO: add logging to tlparse
-    _FX_METADATA_REGISTRY[module_name] = metadata
-
 
 @compatibility(is_backward_compatible=False)
 class NodeSourceAction(Enum):

torch/utils/viz/MemoryViz.js

@@ -806,29 +806,7 @@ function format_frames(frames) {
   }
   const frame_strings = frames
     .filter(frameFilter)
-    .map(f => {
-      let frame_str = `${f.filename}:${f.line}:${f.name}`;
-
-      // Add FX debug information if available
-      if (f.fx_node_op || f.fx_node_name || f.fx_node_target) {
-        const fx_parts = [];
-        if (f.fx_node_name) fx_parts.push(`node=${f.fx_node_name}`);
-        if (f.fx_node_op) fx_parts.push(`op=${f.fx_node_op}`);
-        if (f.fx_node_target) fx_parts.push(`target=${f.fx_node_target}`);
-        frame_str += `\n    >> FX: ${fx_parts.join(', ')}`;
-      }
-
-      if (f.fx_original_trace) {
-        frame_str += `\n    >> Original Model Code:`;
-        const original_lines = f.fx_original_trace.trim().split('\n');
-        // Show all lines of the original trace
-        for (const line of original_lines) {
-          frame_str += `\n       ${line}`;
-        }
-      }
-
-      return frame_str;
-    });
+    .map(f => `${f.filename}:${f.line}:${f.name}`);
   return elideRepeats(frame_strings).join('\n');
 }