Update base for Update on "[Inductor] Mutable custom op pattern matching and safety checker"

TL;DR
TorchInductor now supports pattern matching mutable custom ops directly by unwrapping auto_functionalized wrappers and inserting explicit dependency edges. This enables stable fusion patterns across PyTorch versions.

Problem:
 vLLM has mutable custom ops such as (`rms_norm`, `static_scaled_fp8_quant`) that require pattern matching for [fusion passes](824a3f403f/vllm/compilation/fusion.py (L122-L131)). Currently they pattern match against `auto_functionalized(mutable_op)` wrappers, but vLLM is upgrading to `auto_functionalized_v2` (soon v3) with incompatible semantics that break existing patterns.
`auto_functionalized_v2` decomposes to: view + clone + functional_op + copy_. The specific view operations vary based on which inputs are mutated, making it difficult to write stable patterns that match view+op combinations.


Why current pattern matcher not support the raw custom mutating op ?
Consider this mutable op sequence:

```python
foo_inplace(x)    # Mutates tensor x
bar_out(x, out)   # Uses mutated x, produces out
```

FX Graph Representation:
```python
%x = placeholder()
%out = placeholder()
%foo_result = call_function(foo_inplace, (%x,))
%bar_result = call_function(bar_out, (%x, %out))  # Missing dependency!
```
There is no explicit edge from `foo_inplace` to `bar_out` even though `bar_out` depends on `foo_inplace` mutation. Without explicit edges, pattern matchers cannot reliably detect op sequences or ensure correct execution order.

High level idea:
- Identify mutation ops using operator schemas
- For each mutated tensor, find all storages (including views/aliases) via GraphAliasTracker
- Insert DEP_OP after each mutation
- Redirect later users of aliased storages to depend on DEP_OP

Example:

Custom ops definitions
```python
torch.library.custom_op("mylib::foo_inplace", mutates_args={"x"})
def foo_inplace(x: torch.Tensor) -> None:
    x.add_(1)

torch.library.custom_op("mylib::bar_out", mutates_args={"out"})
def bar_out(x: torch.Tensor, out: torch.Tensor) -> None:
    out.copy_(x + 2)

torch.library.custom_op("mylib::foobar_out", mutates_args={"x", "out"})
def foobar_out(x: torch.Tensor, out: torch.Tensor) -> None:
    x.add_(1)
    out.copy_(x + 2)

# pattern registration
def pattern(x, out):
    foo_inplace(x)
    bar_out(x, out)
    return x, out

def replacement(x, out):
    foobar_out(x, out)
    return x, out
```

Pattern graph after add_implict_edges (used for matching)
```python
graph():
    %x_1 : [num_users=2] = placeholder[target=x_1]
    %out_1 : [num_users=2] = placeholder[target=out_1]
    %foo_inplace : [num_users=1] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%x_1,), kwargs = {})
    %op_for_dependencies : [num_users=2] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%x_1,), kwargs = {writer_token: %foo_inplace})
    %bar_out : [num_users=1] = call_function[target=torch.ops.mylib.bar_out.default](args = (%op_for_dependencies, %out_1), kwargs = {})
    %op_for_dependencies_1 : [num_users=1] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%out_1,), kwargs = {writer_token: %bar_out})
    return (op_for_dependencies, op_for_dependencies_1)
```


Case : mutates a clone of graph input

```python
def f(x, out):
    x = x.clone()
    out = out.clone()
    foo_inplace(x)
    bar_out(x, out)
    return out
```

before mutable custom op pass
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %auto_functionalized_v2 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.foo_inplace.default,), kwargs = {_x_base_index: 0, _all_bases: [%arg0_1]})
    %getitem_1 : [num_users=1] = call_function[target=operator.getitem](args = (%auto_functionalized_v2, 1), kwargs = {})
    %auto_functionalized_v2_1 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.bar_out.default,), kwargs = {x: %getitem_1, _out_base_index: 0, _all_bases: [%arg1_1]})
    %getitem_3 : [num_users=1] = call_function[target=operator.getitem](args = (%auto_functionalized_v2_1, 1), kwargs = {})
    return (getitem_3,)
```

after decompose auto_functionalized
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %as_strided_default_2 : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg0_1, [3], [1], 0), kwargs = {})
    %clone_default_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default_2,), kwargs = {})
    %as_strided_default_3 : [num_users=2] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default_1, [3], [1], 0), kwargs = {})
    %foo_inplace_default : [num_users=0] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%as_strided_default_3,), kwargs = {})
    %as_strided_default : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg1_1, [3], [1], 0), kwargs = {})
    %clone_default : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default,), kwargs = {})
    %as_strided_default_1 : [num_users=2] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default, [3], [1], 0), kwargs = {})
    %bar_out_default : [num_users=0] = call_function[target=torch.ops.mylib.bar_out.default](args = (%as_strided_default_3, %as_strided_default_1), kwargs = {})
    return (as_strided_default_1,)
```

after add_implict_edges
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %as_strided_default_2 : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg0_1, [3], [1], 0), kwargs = {})
    %clone_default_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default_2,), kwargs = {})
    %as_strided_default_3 : [num_users=2] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default_1, [3], [1], 0), kwargs = {})
    %foo_inplace_default : [num_users=1] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%as_strided_default_3,), kwargs = {})
    %op_for_dependencies : [num_users=1] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%as_strided_default_3,), kwargs = {writer_token: %foo_inplace_default})
    %as_strided_default : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg1_1, [3], [1], 0), kwargs = {})
    %clone_default : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default,), kwargs = {})
    %as_strided_default_1 : [num_users=2] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default, [3], [1], 0), kwargs = {})
    %bar_out_default : [num_users=1] = call_function[target=torch.ops.mylib.bar_out.default](args = (%op_for_dependencies, %as_strided_default_1), kwargs = {})
    %op_for_dependencies_1 : [num_users=1] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%as_strided_default_1,), kwargs = {writer_token: %bar_out_default})
    return (op_for_dependencies_1,)
```

after remove_implict_edges (pattern match happened foo_inplace + bar -> foobar_out)
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %as_strided_default_2 : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg0_1, [3], [1], 0), kwargs = {})
    %clone_default_1 : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default_2,), kwargs = {})
    %as_strided_default_3 : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default_1, [3], [1], 0), kwargs = {})
    %as_strided_default : [num_users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg1_1, [3], [1], 0), kwargs = {})
    %clone_default : [num_users=1] = call_function[target=torch.ops.aten.clone.default](args = (%as_strided_default,), kwargs = {})
    %as_strided_default_1 : [num_users=2] = call_function[target=torch.ops.aten.as_strided.default](args = (%clone_default, [3], [1], 0), kwargs = {})
    %foobar_out_default : [num_users=0] = call_function[target=torch.ops.mylib.foobar_out.default](args = (%as_strided_default_3, %as_strided_default_1), kwargs = {})
    return (as_strided_default_1,)
```


Case: multiple writers and readers
```python
def f(
    x: torch.Tensor, y: torch.Tensor, outx: torch.Tensor, outy: torch.Tensor
):
    foo_inplace(x.view(-1))
    foo_inplace(y.view(-1))
    bar_out(x, outx)
    bar_out(y, outy)
    return outx, outy
```

Before mutable custom op pass
```python
graph():
    %arg0_1 : [num_users=2] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=2] = placeholder[target=arg1_1]
    %arg2_1 : [num_users=2] = placeholder[target=arg2_1]
    %arg3_1 : [num_users=2] = placeholder[target=arg3_1]
    %auto_functionalized_v2 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.foo_inplace.default,), kwargs = {_x_base_index: 0, _x_alias: True, _all_bases: [%arg0_1]})
    %getitem_1 : [num_users=2] = call_function[target=operator.getitem](args = (%auto_functionalized_v2, 1), kwargs = {})
    %auto_functionalized_v2_1 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.foo_inplace.default,), kwargs = {_x_base_index: 0, _x_alias: True, _all_bases: [%arg1_1]})
    %getitem_3 : [num_users=2] = call_function[target=operator.getitem](args = (%auto_functionalized_v2_1, 1), kwargs = {})
    %auto_functionalized_v2_2 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.bar_out.default,), kwargs = {x: %getitem_1, _out_base_index: 0, _all_bases: [%arg2_1]})
    %getitem_5 : [num_users=1] = call_function[target=operator.getitem](args = (%auto_functionalized_v2_2, 1), kwargs = {})
    %auto_functionalized_v2_3 : [num_users=1] = call_function[target=torch.ops.higher_order.auto_functionalized_v2](args = (mylib.bar_out.default,), kwargs = {x: %getitem_3, _out_base_index: 0, _all_bases: [%arg3_1]})
    %getitem_7 : [num_users=1] = call_function[target=operator.getitem](args = (%auto_functionalized_v2_3, 1), kwargs = {})
    %copy_ : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg0_1, %getitem_1), kwargs = {})
    %copy__1 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg1_1, %getitem_3), kwargs = {})
    %copy__2 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg2_1, %getitem_5), kwargs = {})
    %copy__3 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg3_1, %getitem_7), kwargs = {})
    return ()
```


after decompose auto_functionalized
```python
graph():
    %arg0_1 : [num_users=3] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=3] = placeholder[target=arg1_1]
    %arg2_1 : [num_users=2] = placeholder[target=arg2_1]
    %arg3_1 : [num_users=2] = placeholder[target=arg3_1]
    %alias_default_1 : [num_users=1] = call_function[target=torch.ops.aten.alias.default](args = (%arg0_1,), kwargs = {})
    %foo_inplace_default_1 : [num_users=0] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%alias_default_1,), kwargs = {})
    %alias_default : [num_users=1] = call_function[target=torch.ops.aten.alias.default](args = (%arg1_1,), kwargs = {})
    %foo_inplace_default : [num_users=0] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%alias_default,), kwargs = {})
    %bar_out_default_1 : [num_users=0] = call_function[target=torch.ops.mylib.bar_out.default](args = (%arg0_1, %arg2_1), kwargs = {})
    %bar_out_default : [num_users=0] = call_function[target=torch.ops.mylib.bar_out.default](args = (%arg1_1, %arg3_1), kwargs = {})
    %copy_ : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg0_1, %arg0_1), kwargs = {})
    %copy__1 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg1_1, %arg1_1), kwargs = {})
    %copy__2 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg2_1, %arg2_1), kwargs = {})
    %copy__3 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg3_1, %arg3_1), kwargs = {})
    return ()
```

after add_implict_edges
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %arg2_1 : [num_users=2] = placeholder[target=arg2_1]
    %arg3_1 : [num_users=2] = placeholder[target=arg3_1]
    %alias_default_1 : [num_users=2] = call_function[target=torch.ops.aten.alias.default](args = (%arg0_1,), kwargs = {})
    %foo_inplace_default_1 : [num_users=1] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%alias_default_1,), kwargs = {})
    %op_for_dependencies : [num_users=2] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%alias_default_1,), kwargs = {writer_token: %foo_inplace_default_1})
    %alias_default : [num_users=2] = call_function[target=torch.ops.aten.alias.default](args = (%arg1_1,), kwargs = {})
    %foo_inplace_default : [num_users=1] = call_function[target=torch.ops.mylib.foo_inplace.default](args = (%alias_default,), kwargs = {})
    %op_for_dependencies_1 : [num_users=2] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%alias_default,), kwargs = {writer_token: %foo_inplace_default})
    %bar_out_default_1 : [num_users=1] = call_function[target=torch.ops.mylib.bar_out.default](args = (%op_for_dependencies, %arg2_1), kwargs = {})
    %op_for_dependencies_2 : [num_users=1] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%arg2_1,), kwargs = {writer_token: %bar_out_default_1})
    %bar_out_default : [num_users=1] = call_function[target=torch.ops.mylib.bar_out.default](args = (%op_for_dependencies_1, %arg3_1), kwargs = {})
    %op_for_dependencies_3 : [num_users=1] = call_function[target=torch.ops.pattern_matcher.op_for_dependencies](args = (%arg3_1,), kwargs = {writer_token: %bar_out_default})
    %copy_ : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%op_for_dependencies, %op_for_dependencies), kwargs = {})
    %copy__1 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%op_for_dependencies_1, %op_for_dependencies_1), kwargs = {})
    %copy__2 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%op_for_dependencies_2, %op_for_dependencies_2), kwargs = {})
    %copy__3 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%op_for_dependencies_3, %op_for_dependencies_3), kwargs = {})
    return ()
```

after remove_implict_edges (pattern match happened foo_inplace + bar -> foobar_out)
```python
graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %arg1_1 : [num_users=1] = placeholder[target=arg1_1]
    %arg2_1 : [num_users=2] = placeholder[target=arg2_1]
    %arg3_1 : [num_users=2] = placeholder[target=arg3_1]
    %alias_default_1 : [num_users=2] = call_function[target=torch.ops.aten.alias.default](args = (%arg0_1,), kwargs = {})
    %alias_default : [num_users=2] = call_function[target=torch.ops.aten.alias.default](args = (%arg1_1,), kwargs = {})
    %foobar_out_default_1 : [num_users=0] = call_function[target=torch.ops.mylib.foobar_out.default](args = (%alias_default_1, %arg2_1), kwargs = {})
    %foobar_out_default : [num_users=0] = call_function[target=torch.ops.mylib.foobar_out.default](args = (%alias_default, %arg3_1), kwargs = {})
    %copy_ : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%alias_default_1, %alias_default_1), kwargs = {})
    %copy__1 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%alias_default, %alias_default), kwargs = {})
    %copy__2 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg2_1, %arg2_1), kwargs = {})
    %copy__3 : [num_users=0] = call_function[target=torch.ops.aten.copy_.default](args = (%arg3_1, %arg3_1), kwargs = {})
    return ()
```

cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos choijon5

[ghstack-poisoned]

.ci/docker/almalinux/build.sh

@@ -36,7 +36,11 @@ case ${DOCKER_TAG_PREFIX} in
     ;;
   rocm*)
     BASE_TARGET=rocm
-    PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx950;gfx1150;gfx1151"
+    PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+    # add gfx950, gfx115x conditionally starting in ROCm 7.0
+    if [[ "$ROCM_VERSION" == *"7.0"* ]]; then
+        PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950;gfx1150;gfx1151"
+    fi
     EXTRA_BUILD_ARGS="${EXTRA_BUILD_ARGS} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}"
     ;;
   *)

.ci/docker/build.sh

@@ -168,18 +168,6 @@ case "$tag" in
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.11-clang12)
-    ANACONDA_PYTHON_VERSION=3.11
-    CLANG_VERSION=12
-    VISION=no
-    TRITON=no
-    ;;
-  pytorch-linux-jammy-py3.12-clang12)
-    ANACONDA_PYTHON_VERSION=3.12
-    CLANG_VERSION=12
-    VISION=no
-    TRITON=no
-    ;;
   pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
       ANACONDA_PYTHON_VERSION=3.10
@@ -207,9 +195,9 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-noble-xpu-n-py3 | pytorch-linux-noble-xpu-n-py3-inductor-benchmarks)
+  pytorch-linux-jammy-xpu-n-py3 | pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=13
+    GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.2
     NINJA_VERSION=1.9.0
@@ -260,12 +248,6 @@ case "$tag" in
     HALIDE=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-py3.12-pallas)
-    CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.12
-    GCC_VERSION=11
-    PALLAS=yes
-    ;;
   pytorch-linux-jammy-py3.12-triton-cpu)
     CUDA_VERSION=12.6
     ANACONDA_PYTHON_VERSION=3.12
@@ -387,7 +369,6 @@ docker build \
        --build-arg "INDUCTOR_BENCHMARKS=${INDUCTOR_BENCHMARKS}" \
        --build-arg "EXECUTORCH=${EXECUTORCH}" \
        --build-arg "HALIDE=${HALIDE}" \
-       --build-arg "PALLAS=${PALLAS}" \
        --build-arg "XPU_VERSION=${XPU_VERSION}" \
        --build-arg "UNINSTALL_DILL=${UNINSTALL_DILL}" \
        --build-arg "ACL=${ACL:-}" \

.ci/docker/ci_commit_pins/jax.txt

@@ -1 +0,0 @@
-0.8.0

.ci/docker/common/install_jax.sh

@@ -1,40 +0,0 @@
-#!/bin/bash
-
-set -ex
-
-source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
-
-# Get the pinned JAX version (same for all CUDA versions)
-JAX_VERSION=$(get_pinned_commit /ci_commit_pins/jax)
-
-function install_jax_12() {
-  echo "Installing JAX ${JAX_VERSION} with CUDA 12 support"
-  pip_install "jax[cuda12]==${JAX_VERSION}" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
-
-  # Verify installation
-  python -c "import jax"  # check for errors
-  echo "JAX ${JAX_VERSION} installation completed successfully for CUDA 12"
-}
-
-function install_jax_13() {
-  echo "Installing JAX ${JAX_VERSION} with CUDA 13 support"
-  pip_install "jax[cuda13]==${JAX_VERSION}" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
-
-  # Verify installation
-  python -c "import jax"  # check for errors
-  echo "JAX ${JAX_VERSION} installation completed successfully for CUDA 13"
-}
-
-# idiomatic parameter and option handling in sh
-while test $# -gt 0
-do
-    case "$1" in
-    12.4|12.6|12.6.*|12.8|12.8.*|12.9|12.9.*) install_jax_12;
-        ;;
-    13.0|13.0.*) install_jax_13;
-        ;;
-    *) echo "bad argument $1"; exit 1
-        ;;
-    esac
-    shift
-done

.ci/docker/common/install_xpu.sh

@@ -9,7 +9,7 @@ set -xe
 
 function install_ubuntu() {
     . /etc/os-release
-    if [[ ! " jammy noble " =~ " ${VERSION_CODENAME} " ]]; then
+    if [[ ! " jammy " =~ " ${VERSION_CODENAME} " ]]; then
         echo "Ubuntu version ${VERSION_CODENAME} not supported"
         exit
     fi
@@ -35,24 +35,25 @@ function install_ubuntu() {
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
 
-    # Compute and Media Runtimes
-    if [[ " ${VERSION_CODENAME} " =~ " noble " ]]; then
+    if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
+        # Compute and Media Runtimes
         apt-get install -y \
-            intel-opencl-icd libze-intel-gpu1 libze1 \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+            intel-opencl-icd intel-level-zero-gpu level-zero \
+            intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
+            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
             libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
-    else # jammy
+            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
+        # Development Packages
+        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
+    else # rolling driver
         apt-get install -y \
             intel-opencl-icd libze-intel-gpu1 libze1 \
             intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
             libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
             libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
             mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
+        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
     fi
-    # Development Packages
-    apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
 
     # Install Intel Support Packages
     apt-get install -y ${XPU_PACKAGES}
@@ -65,7 +66,7 @@ function install_ubuntu() {
 function install_rhel() {
     . /etc/os-release
     if [[ "${ID}" == "rhel" ]]; then
-        if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
+        if [[ ! " 8.8 8.9 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
             echo "RHEL version ${VERSION_ID} not supported"
             exit
         fi
@@ -146,7 +147,7 @@ function install_sles() {
 XPU_DRIVER_VERSION=""
 if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
     # Use GPU driver LTS releases
-    XPU_DRIVER_VERSION="/lts/2523"
+    XPU_DRIVER_VERSION="/lts/2350"
 fi
 
 # Default use Intel® oneAPI Deep Learning Essentials 2025.1

.ci/docker/libtorch/build.sh

@@ -49,7 +49,11 @@ case ${DOCKER_TAG_PREFIX} in
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
-        PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx950;gfx1150;gfx1151"
+        PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+        # add gfx950, gfx115x conditionally starting in ROCm 7.0
+        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
+            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950;gfx1150;gfx1151"
+        fi
         DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg ROCM_VERSION=${GPU_ARCH_VERSION}"
         ;;
     *)

.ci/docker/manywheel/build.sh

@@ -87,7 +87,11 @@ case ${image} in
         MANY_LINUX_VERSION="2_28"
         DEVTOOLSET_VERSION="11"
         GPU_IMAGE=rocm/dev-almalinux-8:${GPU_ARCH_VERSION}-complete
-        PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx950;gfx1150;gfx1151"
+        PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+        # add gfx950, gfx115x conditionally starting in ROCm 7.0
+        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
+            PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950;gfx1150;gfx1151"
+        fi
         DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}"
         ;;
     manylinux2_28-builder:xpu)

.ci/docker/ubuntu/Dockerfile

@@ -143,15 +143,6 @@ COPY ci_commit_pins/halide.txt halide.txt
 RUN if [ -n "${HALIDE}" ]; then bash ./install_halide.sh; fi
 RUN rm install_halide.sh common_utils.sh halide.txt
 
-ARG PALLAS
-ARG CUDA_VERSION
-# Install JAX with CUDA support (for Pallas)
-COPY ./common/install_jax.sh install_jax.sh
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/jax.txt /ci_commit_pins/jax.txt
-RUN if [ -n "${PALLAS}" ]; then bash ./install_jax.sh ${CUDA_VERSION}; fi
-RUN rm -f install_jax.sh common_utils.sh /ci_commit_pins/jax.txt
-
 ARG ONNX
 # Install ONNX dependencies
 COPY ./common/install_onnx.sh ./common/common_utils.sh ./

.ci/lumen_cli/cli/lib/common/cli_helper.py

@@ -8,11 +8,9 @@ from abc import ABC, abstractmethod
 
 
 try:
-    from collections.abc import Callable  # Python 3.11+
-    from typing import Any, Required, TypedDict
+    from typing import Any, Callable, Required, TypedDict  # Python 3.11+
 except ImportError:
-    from collections.abc import Callable
-    from typing import Any, TypedDict
+    from typing import Any, Callable, TypedDict
 
     from typing_extensions import Required  # Fallback for Python <3.11
 

.ci/pytorch/build.sh

@@ -168,16 +168,14 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   # shellcheck disable=SC1091
   source /opt/intel/oneapi/compiler/latest/env/vars.sh
   # shellcheck disable=SC1091
-  source /opt/intel/oneapi/umf/latest/env/vars.sh
-  # shellcheck disable=SC1091
   source /opt/intel/oneapi/ccl/latest/env/vars.sh
   # shellcheck disable=SC1091
   source /opt/intel/oneapi/mpi/latest/env/vars.sh
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/pti/latest/env/vars.sh
   # Enable XCCL build
   export USE_XCCL=1
   export USE_MPI=0
+  # XPU kineto feature dependencies are not fully ready, disable kineto build as temp WA
+  export USE_KINETO=0
   export TORCH_XPU_ARCH_LIST=pvc
 fi
 

.ci/pytorch/test.sh

@@ -208,8 +208,6 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   source /opt/intel/oneapi/ccl/latest/env/vars.sh
   # shellcheck disable=SC1091
   source /opt/intel/oneapi/mpi/latest/env/vars.sh
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/pti/latest/env/vars.sh
   # Check XPU status before testing
   timeout 30 xpu-smi discovery || true
 fi
@@ -826,11 +824,6 @@ test_inductor_halide() {
   assert_git_not_dirty
 }
 
-test_inductor_pallas() {
-  python test/run_test.py --include inductor/test_pallas.py --verbose
-  assert_git_not_dirty
-}
-
 test_inductor_triton_cpu() {
   python test/run_test.py --include inductor/test_triton_cpu_backend.py inductor/test_torchinductor_strided_blocks.py --verbose
   assert_git_not_dirty
@@ -1731,8 +1724,6 @@ elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
   test_inductor_halide
-elif [[ "${TEST_CONFIG}" == *inductor-pallas* ]]; then
-  test_inductor_pallas
 elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-ccb801b88af136454798b945175c4c87e636ac33
+cfbc5c2f1c798991715a6b06bb3ce46478c4487c

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-e4d25697f9dc5eedaf8f0a5bf085c62c5455a53a
+c8b09f5f77d6bf6fb7ed7a9aa83e5d8156b3a5e9

.github/labeler.yml

@@ -138,8 +138,7 @@
 - test/test_matmul_cuda.py
 - test/test_scaled_matmul_cuda.py
 - test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/*Blas.cpp
-- aten/src/ATen/cuda/CUDA*Blas.*
+- aten/src/ATen/native/cuda/Blas.cpp
 - torch/**/*cublas*
 - torch/_inductor/kernel/mm.py
 - test/inductor/test_max_autotune.py
@@ -149,8 +148,7 @@
 - test/test_matmul_cuda.py
 - test/test_scaled_matmul_cuda.py
 - test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/*Blas.cpp
-- aten/src/ATen/cuda/CUDA*Blas.*
+- aten/src/ATen/native/cuda/Blas.cpp
 - torch/**/*cublas*
 - torch/_inductor/kernel/mm.py
 - test/inductor/test_max_autotune.py
@@ -160,8 +158,7 @@
 - test/test_matmul_cuda.py
 - test/test_scaled_matmul_cuda.py
 - test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/*Blas.cpp
-- aten/src/ATen/cuda/CUDA*Blas.*
+- aten/src/ATen/native/cuda/Blas.cpp
 - torch/_inductor/kernel/mm.py
 - test/inductor/test_max_autotune.py
 - third_party/fbgemm

.github/nitpicks.yml

@@ -10,4 +10,3 @@
   pathFilter:
     - 'torch/csrc/inductor/aoti_torch/c/*'
     - 'torch/csrc/inductor/aoti_torch/generated/*'
-    - 'torch/csrc/stable/c/*'

.github/pytorch-probot.yml

@@ -2,8 +2,8 @@ tracking_issue: 24422
 ciflow_tracking_issue: 64124
 ciflow_push_tags:
 - ciflow/b200
-- ciflow/b200-distributed
 - ciflow/b200-symm-mem
+- ciflow/b200-distributed
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
@@ -22,8 +22,6 @@ ciflow_push_tags:
 - ciflow/inductor-perf-test-nightly-xpu
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
-- ciflow/inductor-rocm-mi200
-- ciflow/inductor-rocm-mi300
 - ciflow/linux-aarch64
 - ciflow/mps
 - ciflow/nightly
@@ -35,13 +33,11 @@ ciflow_push_tags:
 - ciflow/quantization-periodic
 - ciflow/riscv64
 - ciflow/rocm
-- ciflow/rocm-mi200
 - ciflow/rocm-mi300
 - ciflow/rocm-mi355
 - ciflow/rocm-navi31
 - ciflow/s390
 - ciflow/slow
-- ciflow/slow-rocm-mi200
 - ciflow/torchbench
 - ciflow/triton_binaries
 - ciflow/trunk

.github/scripts/delete_old_branches.py

@@ -1,11 +1,10 @@
 # Delete old branches
 import os
 import re
-from collections.abc import Callable
 from datetime import datetime
 from functools import lru_cache
 from pathlib import Path
-from typing import Any
+from typing import Any, Callable
 
 from github_utils import gh_fetch_json_dict, gh_graphql
 from gitutils import GitRepo

.github/scripts/filter_test_configs.py

@@ -8,11 +8,10 @@ import re
 import subprocess
 import sys
 import warnings
-from collections.abc import Callable
 from enum import Enum
 from functools import cache
 from logging import info
-from typing import Any, Optional
+from typing import Any, Callable, Optional
 from urllib.request import Request, urlopen
 
 import yaml

.github/scripts/get_workflow_job_id.py

@@ -11,8 +11,7 @@ import sys
 import time
 import urllib
 import urllib.parse
-from collections.abc import Callable
-from typing import Any, Optional
+from typing import Any, Callable, Optional
 from urllib.request import Request, urlopen
 
 

.github/scripts/github_utils.py

@@ -3,9 +3,8 @@
 import json
 import os
 import warnings
-from collections.abc import Callable
 from dataclasses import dataclass
-from typing import Any, cast, Optional, Union
+from typing import Any, Callable, cast, Optional, Union
 from urllib.error import HTTPError
 from urllib.parse import quote
 from urllib.request import Request, urlopen

.github/scripts/gitutils.py

@@ -4,10 +4,10 @@ import os
 import re
 import tempfile
 from collections import defaultdict
-from collections.abc import Callable, Iterator
+from collections.abc import Iterator
 from datetime import datetime
 from functools import wraps
-from typing import Any, cast, Optional, TypeVar, Union
+from typing import Any, Callable, cast, Optional, TypeVar, Union
 
 
 T = TypeVar("T")

.github/scripts/trymerge.py

@@ -17,12 +17,12 @@ import re
 import time
 import urllib.parse
 from collections import defaultdict
-from collections.abc import Callable, Iterable
+from collections.abc import Iterable
 from dataclasses import dataclass
 from functools import cache
 from pathlib import Path
 from re import Pattern
-from typing import Any, cast, NamedTuple, Optional
+from typing import Any, Callable, cast, NamedTuple, Optional
 from warnings import warn
 
 import yaml

.github/workflows/docker-builds.yml

@@ -56,8 +56,6 @@ jobs:
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.10-clang12,
-          pytorch-linux-jammy-py3.11-clang12,
-          pytorch-linux-jammy-py3.12-clang12,
           pytorch-linux-jammy-py3.13-clang12,
           pytorch-linux-jammy-py3.14-clang12,
           pytorch-linux-jammy-rocm-n-py3,
@@ -67,10 +65,9 @@ jobs:
           pytorch-linux-jammy-py3.10-gcc11,
           pytorch-linux-jammy-py3-gcc11-inductor-benchmarks,
           pytorch-linux-jammy-py3.12-halide,
-          pytorch-linux-jammy-cuda12.8-py3.12-pallas,
           pytorch-linux-jammy-xpu-n-1-py3,
-          pytorch-linux-noble-xpu-n-py3,
-          pytorch-linux-noble-xpu-n-py3-inductor-benchmarks,
+          pytorch-linux-jammy-xpu-n-py3,
+          pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,

.github/workflows/inductor-perf-test-nightly-xpu.yml

@@ -83,8 +83,8 @@ jobs:
     needs: get-label-type
     with:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-noble-xpu-n-py3.10
-      docker-image-name: ci-image:pytorch-linux-noble-xpu-n-py3-inductor-benchmarks
+      build-environment: linux-jammy-xpu-n-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks
       runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
@@ -117,7 +117,7 @@ jobs:
     uses: ./.github/workflows/_xpu-test.yml
     needs: xpu-n-py3_10-inductor-benchmark-build
     with:
-      build-environment: linux-noble-xpu-n-py3.10
+      build-environment: linux-jammy-xpu-n-py3.10
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
       docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
@@ -137,7 +137,7 @@ jobs:
     uses: ./.github/workflows/_xpu-test.yml
     needs: xpu-n-py3_10-inductor-benchmark-build
     with:
-      build-environment: linux-noble-xpu-n-py3.10
+      build-environment: linux-jammy-xpu-n-py3.10
       dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
       docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
       test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}

.github/workflows/inductor-rocm-mi300.yml

@@ -7,7 +7,6 @@ on:
       - release/*
     tags:
       - ciflow/inductor-rocm/*
-      - ciflow/inductor-rocm-mi300/*
   workflow_dispatch:
 
 concurrency:

.github/workflows/inductor-rocm.yml

@@ -7,7 +7,7 @@ on:
     branches:
       - release/*
     tags:
-      - ciflow/inductor-rocm-mi200/*
+      - ciflow/inductor-rocm/*
   workflow_dispatch:
 
 concurrency:

.github/workflows/inductor-unittest.yml

@@ -81,32 +81,6 @@ jobs:
       test-matrix: ${{ needs.inductor-halide-build.outputs.test-matrix }}
     secrets: inherit
 
-  inductor-pallas-build:
-    name: inductor-pallas-build
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.12-gcc11
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-py3.12-pallas
-      cuda-arch-list: '8.9'
-      runner: linux.8xlarge.memory
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      test-matrix: |
-        { include: [
-          { config: "inductor-pallas", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g5.12xlarge.nvidia.gpu" },
-        ]}
-    secrets: inherit
-
-  inductor-pallas-test:
-    name: inductor-pallas-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: inductor-pallas-build
-    with:
-      build-environment: linux-jammy-py3.12-gcc11
-      docker-image: ${{ needs.inductor-pallas-build.outputs.docker-image }}
-      test-matrix: ${{ needs.inductor-pallas-build.outputs.test-matrix }}
-    secrets: inherit
-
   inductor-triton-cpu-build:
     name: inductor-triton-cpu-build
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/periodic-rocm-mi200.yml

@@ -11,6 +11,7 @@ on:
     - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
   push:
     tags:
+      - ciflow/periodic/*
       - ciflow/periodic-rocm-mi200/*
     branches:
       - release/*

.github/workflows/periodic-rocm-mi300.yml

@@ -11,7 +11,6 @@ on:
     - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
   push:
     tags:
-      - ciflow/periodic/*
       - ciflow/periodic-rocm-mi300/*
     branches:
       - release/*

.github/workflows/pull.yml

@@ -342,16 +342,16 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc9-inductor-build.outputs.test-matrix }}
     secrets: inherit
 
-  linux-noble-xpu-n-py3_10-build:
-    name: linux-noble-xpu-n-py3.10
+  linux-jammy-xpu-n-py3_10-build:
+    name: linux-jammy-xpu-n-py3.10
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       # This should sync with the build in xpu.yml but xpu uses a larger runner
       # sync-tag: linux-xpu-n-build
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
-      build-environment: linux-noble-xpu-n-py3.10
-      docker-image-name: ci-image:pytorch-linux-noble-xpu-n-py3
+      build-environment: linux-jammy-xpu-n-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 4, runner: "linux.idc.xpu" },

.github/workflows/rocm-mi300.yml

@@ -6,7 +6,6 @@ on:
       - main
       - release/*
     tags:
-      - ciflow/rocm/*
       - ciflow/rocm-mi300/*
   workflow_dispatch:
   schedule:

.github/workflows/rocm.yml

@@ -5,7 +5,7 @@ on:
     branches:
       - release/*
     tags:
-      - ciflow/rocm-mi200/*
+      - ciflow/rocm/*
   workflow_dispatch:
   schedule:
     - cron: 29 8 * * *  # about 1:29am PDT

.github/workflows/slow-rocm-mi200.yml

@@ -1,81 +0,0 @@
-# This workflow is dedicated to host slow jobs that are run only periodically because
-# they are too slow to run in every commit.  The list of slow tests can be found in
-# https://github.com/pytorch/test-infra/blob/generated-stats/stats/slow-tests.json
-name: slow-rocm-mi200
-
-on:
-  push:
-    branches:
-      - release/*
-    tags:
-      - ciflow/slow/*
-      - ciflow/slow-rocm-mi200/*
-  schedule:
-    - cron: 0 */3 * * *
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  llm-td:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/llm_td_retrieval.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  target-determination:
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    needs: llm-td
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "slow", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.2", owners: ["module:rocm"] },
-          { config: "slow", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.2", owners: ["module:rocm"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/slow.yml

@@ -105,6 +105,36 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-py3_10-clang12-build.outputs.test-matrix }}
     secrets: inherit
 
+  linux-jammy-rocm-py3_10-build:
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      test-matrix: |
+        { include: [
+          { config: "slow", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.2", owners: ["module:rocm"] },
+          { config: "slow", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.2", owners: ["module:rocm"] },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit
+
   linux-jammy-py3_10-clang18-asan-build:
     name: linux-jammy-py3.10-clang18-asan
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/upload-test-stats.yml

@@ -11,16 +11,15 @@ on:
       - inductor
       - unstable
       - slow
-      - slow-rocm-mi200
       - unstable-periodic
       - inductor-periodic
-      - rocm-mi200
+      - rocm
       - rocm-mi300
       - rocm-mi355
       - inductor-micro-benchmark
       - inductor-micro-benchmark-x86
       - inductor-cu124
-      - inductor-rocm-mi200
+      - inductor-rocm
       - inductor-rocm-mi300
       - mac-mps
       - linux-aarch64

.github/workflows/xpu.yml

@@ -47,15 +47,15 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-noble-xpu-n-py3_10-build:
-    name: linux-noble-xpu-n-py3.10
+  linux-jammy-xpu-n-py3_10-build:
+    name: linux-jammy-xpu-n-py3.10
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
       sync-tag: linux-xpu-n-build
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
-      build-environment: linux-noble-xpu-n-py3.10
-      docker-image-name: ci-image:pytorch-linux-noble-xpu-n-py3
+      build-environment: linux-jammy-xpu-n-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3
       runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
@@ -74,17 +74,17 @@ jobs:
         ]}
     secrets: inherit
 
-  linux-noble-xpu-n-py3_10-test:
-    name: linux-noble-xpu-n-py3.10
+  linux-jammy-xpu-n-py3_10-test:
+    name: linux-jammy-xpu-n-py3.10
     uses: ./.github/workflows/_xpu-test.yml
-    needs: linux-noble-xpu-n-py3_10-build
+    needs: linux-jammy-xpu-n-py3_10-build
     permissions:
       id-token: write
       contents: read
     with:
-      build-environment: linux-noble-xpu-n-py3.10
-      docker-image: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-noble-xpu-n-py3_10-build.outputs.test-matrix }}
+      build-environment: linux-jammy-xpu-n-py3.10
+      docker-image: ${{ needs.linux-jammy-xpu-n-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-xpu-n-py3_10-build.outputs.test-matrix }}
     secrets: inherit
 
   windows-xpu-n-1-build:

.lintrunner.toml

@@ -143,8 +143,7 @@ init_command = [
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
     'numpy==1.26.4 ; python_version >= "3.10" and python_version <= "3.11"',
-    'numpy==2.1.0 ; python_version >= "3.12" and python_version <= "3.13"',
-    'numpy==2.3.4 ; python_version >= "3.14"',
+    'numpy==2.1.0 ; python_version >= "3.12"',
     'expecttest==0.3.0',
     'pyrefly==0.36.2',
     'sympy==1.13.3',
@@ -1402,7 +1401,7 @@ init_command = [
     '--dry-run={{DRYRUN}}',
     'usort==1.0.8.post1',
     'isort==6.0.1',
-    'ruff==0.14.4',  # sync with RUFF
+    'ruff==0.13.1',  # sync with RUFF
 ]
 is_formatter = true
 
@@ -1537,7 +1536,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'ruff==0.14.4',  # sync with PYFMT
+    'ruff==0.13.1',  # sync with PYFMT
 ]
 is_formatter = true
 

CODEOWNERS

@@ -210,12 +210,8 @@ torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
 /test/inductor/test_flex_attention.py @drisspg
 /test/inductor/test_flex_decoding.py @drisspg
 
-# Low Precision & Grouped GEMMs
+# Low Precision GEMMs
 /aten/src/ATen/native/cuda/Blas.cpp @drisspg @slayton58
-/aten/src/ATen/native/cuda/GroupedBlas.cpp @drisspg @slayton58
-/aten/src/ATen/native/cuda/ScaledBlas.cpp @drisspg @slayton58
 /aten/src/ATen/cuda/CUDABlas.cpp @drisspg @slayton58
 /aten/src/ATen/cuda/CUDABlas.h @drisspg @slayton58
-/aten/src/ATen/cuda/CUDAScaledBlas.cpp @drisspg @slayton58
-/aten/src/ATen/cuda/CUDAScaledBlas.h @drisspg @slayton58
 /test/test_scaled_matmul_cuda.py @drisspg @slayton58

aten/src/ATen/Context.h

@@ -174,12 +174,6 @@ class TORCH_API Context {
   static long versionCuDNN() {
     return detail::getCUDAHooks().versionCuDNN();
   }
-  static long versionRuntimeCuDNN() {
-    return detail::getCUDAHooks().versionRuntimeCuDNN();
-  }
-  static long versionCuDNNFrontend() {
-    return detail::getCUDAHooks().versionCuDNNFrontend();
-  }
   static bool hasCuSOLVER() {
     return detail::getCUDAHooks().hasCuSOLVER();
   }

aten/src/ATen/DeviceAccelerator.h

@@ -94,11 +94,6 @@ TORCH_API inline void resetPeakStats(c10::DeviceIndex device_index) {
   at::getDeviceAllocator(device_type)->resetPeakStats(device_index);
 }
 
-TORCH_API inline std::pair<size_t, size_t> getMemoryInfo(
-    c10::DeviceIndex device_index) {
-  const auto device_type = getAccelerator(true).value();
-  return at::getDeviceAllocator(device_type)->getMemoryInfo(device_index);
-}
 } // namespace at::accelerator
 
 namespace at {

aten/src/ATen/Dispatch.h

@@ -6,7 +6,6 @@
 #include <c10/util/Half.h>
 #include <c10/util/Metaprogramming.h>
 #include <c10/util/complex.h>
-#include <torch/headeronly/core/Dispatch.h>
 
 #ifdef __CUDACC__
 #include <cuda.h> // For CUDA_VERSION
@@ -62,9 +61,12 @@ TORCH_API void record_kernel_function_dtype(std::string name);
     }                                                 \
   } while (0)
 
-#define AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, HINT, ...) \
-  THO_PRIVATE_CASE_TYPE_USING_HINT_TMPL(                      \
-      AT_PRIVATE_CHECK_SELECTIVE_BUILD, enum_type, HINT, __VA_ARGS__)
+#define AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, HINT, ...)                 \
+  case enum_type: {                                                           \
+    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                              \
+    using HINT [[maybe_unused]] = c10::impl::ScalarTypeToCPPTypeT<enum_type>; \
+    return __VA_ARGS__();                                                     \
+  }
 
 #define AT_DISPATCH_CASE(enum_type, ...) \
   AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, scalar_t, __VA_ARGS__)
@@ -93,6 +95,14 @@ TORCH_API void record_kernel_function_dtype(std::string name);
     return __VA_ARGS__();                                                   \
   }
 
+namespace detail {
+
+inline at::ScalarType scalar_type(at::ScalarType s) {
+  return s;
+}
+
+} // namespace detail
+
 // The AT_DISPATCH_* family of macros provides the ability to
 // conveniently generate specializations of a kernel over all of the
 // dtypes we care about in PyTorch.  We call it "dispatch" because
@@ -180,13 +190,27 @@ TORCH_API void record_kernel_function_dtype(std::string name);
 // but we're just being safe (and it doesn't hurt.)  Note we must
 // use it to shut up warnings about unused store.
 
-#define AT_DISPATCH_SWITCH(TYPE, NAME, ...) \
-  THO_DISPATCH_SWITCH_TMPL(                 \
-      RECORD_KERNEL_FUNCTION_DTYPE,         \
-      TORCH_CHECK_NOT_IMPLEMENTED,          \
-      TYPE,                                 \
-      NAME,                                 \
-      __VA_ARGS__)
+#define AT_DISPATCH_SWITCH(TYPE, NAME, ...)                                 \
+  [&] {                                                                     \
+    const auto& the_type = TYPE;                                            \
+    constexpr const char* at_dispatch_name = NAME;                          \
+    /* don't use TYPE again in case it is an expensive or side-effect op */ \
+    at::ScalarType _st = ::detail::scalar_type(the_type);                   \
+    RECORD_KERNEL_FUNCTION_DTYPE(at_dispatch_name, _st);                    \
+    C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")             \
+    switch (_st) {                                                          \
+      __VA_ARGS__                                                           \
+      default:                                                              \
+        TORCH_CHECK_NOT_IMPLEMENTED(                                        \
+            false,                                                          \
+            '"',                                                            \
+            at_dispatch_name,                                               \
+            "\" not implemented for '",                                     \
+            toString(_st),                                                  \
+            "'");                                                           \
+    }                                                                       \
+    C10_DIAGNOSTIC_POP()                                                    \
+  }()
 
 #define AT_DISPATCH_CASE_FLOATING_TYPES(...)            \
   AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \

aten/src/ATen/Dispatch_v2.h

@@ -1,8 +1,3 @@
-#pragma once
-
-#include <torch/headeronly/core/Dispatch_v2.h>
-
-// Get AT_DISPATCH_SWITCH and AT_DISPATCH_CASE:
 #include <ATen/Dispatch.h>
 
 // This is a new implementation of the AT_DISPATCH macro family from
@@ -79,19 +74,41 @@
 // macro expansion occurs, mediated with AT_EXPAND and AT_GUARD.  I mostly
 // relied on GPT4 to help me get it right.
 
+// Public API macros
+
 // See documentation above
 #define AT_DISPATCH_V2(TYPE, NAME, BODY, ...) \
-  THO_DISPATCH_V2_TMPL(                       \
-      AT_DISPATCH_SWITCH,                     \
-      AT_DISPATCH_CASE,                       \
-      TYPE,                                   \
-      NAME,                                   \
-      AT_WRAP(BODY),                          \
-      __VA_ARGS__)
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_AP_VAR(AT_WRAP(BODY), TYPE, __VA_ARGS__))
+
+// This macro lets you pass an arbitrary expression that may contain internal
+// commas to another macro without having the commas causing the expression
+// to be interpreted as being multiple arguments
+#define AT_WRAP(...) __VA_ARGS__
+
+#define AT_FLOAT8_TYPES                                          \
+  c10::kFloat8_e5m2, c10::kFloat8_e5m2fnuz, c10::kFloat8_e4m3fn, \
+      c10::kFloat8_e4m3fnuz, c10::kFloat8_e8m0fnu
+
+#define AT_INTEGRAL_TYPES \
+  c10::kByte, c10::kChar, c10::kInt, c10::kLong, c10::kShort
+#define AT_FLOATING_TYPES c10::kDouble, c10::kFloat
+#define AT_BAREBONES_UNSIGNED_TYPES c10::kUInt16, c10::kUInt32, c10::kUInt64
+#define AT_INTEGRAL_TYPES_V2 \
+  AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES)
+#define AT_COMPLEX_TYPES c10::kComplexDouble, c10::kComplexFloat
+#define AT_QINT_TYPES c10::kQInt8, c10::kQUInt8, c10::kQInt32
+// NB: not *actually* all types
+#define AT_ALL_TYPES AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_FLOATING_TYPES)
+#define AT_ALL_TYPES_AND_COMPLEX \
+  AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_COMPLEX_TYPES)
+
+// Helper macros
 
-// Unused helper macros, kept for BC:
 #define AT_AP_VAR(N, T, ...) \
   AT_EXPAND(AT_CONCAT(AT_AP, AT_NUM_ARGS(__VA_ARGS__))(AT_WRAP(N), __VA_ARGS__))
+#define AT_CONCAT(a, b) AT_CONCAT_AUX(a, b)
+#define AT_CONCAT_AUX(a, b) a##b
+#define AT_EXPAND(X) X
 
 // Ensure we never have too many scalar types for the expansion here to
 // support.  To bump this, you must regenerate the macros below.
@@ -102,6 +119,12 @@ static_assert(static_cast<int>(c10::ScalarType::NumOptions) < 60);
 
 num_args = 60
 
+nums = ', '.join(str(i) for i in reversed(range(num_args+1)))
+args = ', '.join(f'_{i}' for i in range(1, num_args+1))
+
+print(f'#define AT_NUM_ARGS(...) AT_EXPAND(AT_NUM_ARGS_AUX(__VA_ARGS__, {nums}))')
+print(f'#define AT_NUM_ARGS_AUX({args}, N, ...) N')
+
 for i in range(1, num_args+1):
     args = ', '.join(f'_{i}' for i in range(1, i+1))
     cases = ' '.join([f'AT_DISPATCH_CASE(_{j}, N)' for j in range(1, i+1)])
@@ -112,6 +135,8 @@ for i in range(1, num_args+1):
 // Begin generated code
 // clang-format off
 
+#define AT_NUM_ARGS(...) AT_EXPAND(AT_NUM_ARGS_AUX(__VA_ARGS__, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0))
+#define AT_NUM_ARGS_AUX(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, N, ...) N
 #define AT_AP1(N, _1) AT_DISPATCH_CASE(_1, N)
 #define AT_AP2(N, _1, _2) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N)
 #define AT_AP3(N, _1, _2, _3) AT_DISPATCH_CASE(_1, N) AT_DISPATCH_CASE(_2, N) AT_DISPATCH_CASE(_3, N)

aten/src/ATen/core/CachingHostAllocator.h

@@ -226,8 +226,8 @@ template <
     typename B = HostBlock<S>>
 struct CachingHostAllocatorImpl {
   virtual ~CachingHostAllocatorImpl() {
-    if (active_) {
-      active_ = false;
+    active_ = false;
+    if (pinned_use_background_threads()) {
       getBackgroundThreadPool()->waitWorkComplete();
     }
   }
@@ -260,7 +260,6 @@ struct CachingHostAllocatorImpl {
     if (pinned_use_background_threads()) {
       // Launch the background thread and process events in a loop.
       static bool background_thread_flag [[maybe_unused]] = [this] {
-        active_ = true;
         getBackgroundThreadPool()->run([&]() {
           while (active_) {
             process_events();
@@ -684,9 +683,9 @@ struct CachingHostAllocatorImpl {
   alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
   std::deque<std::pair<E, B*>> events_; // event queue paired with block
 
-  // Indicates whether the event-processing thread pool is active.
+  // Indicates whether the object is active.
   // Set to false in the destructor to signal background threads to stop.
-  std::atomic<bool> active_{false};
+  std::atomic<bool> active_{true};
 protected:
   alignas(hardware_destructive_interference_size) HostStatsStaged stats_;
 };

aten/src/ATen/cuda/detail/CUDAHooks.cpp

@@ -21,7 +21,6 @@
 
 #if AT_CUDNN_ENABLED()
 #include <ATen/cudnn/cudnn-wrapper.h>
-#include <cudnn_frontend.h>
 #endif
 
 #if AT_MAGMA_ENABLED()
@@ -352,26 +351,6 @@ long CUDAHooks::versionCuDNN() const {
 #endif
 }
 
-long CUDAHooks::versionRuntimeCuDNN() const {
-#if AT_CUDNN_ENABLED()
-#ifndef USE_STATIC_CUDNN
-  return cudnnGetVersion();
-#else
-  return CUDNN_VERSION;
-#endif
-#else
-  TORCH_CHECK(false, "Cannot query CuDNN version if ATen_cuda is not built with CuDNN");
-#endif
-}
-
-long CUDAHooks::versionCuDNNFrontend() const {
-#if AT_CUDNN_ENABLED()
-  return CUDNN_FRONTEND_VERSION;
-#else
-  TORCH_CHECK(false, "Cannot query CuDNN Frontend version if ATen_cuda is not built with CuDNN");
-#endif
-}
-
 long CUDAHooks::versionMIOpen() const {
 #if AT_ROCM_ENABLED()
   return MIOPEN_VERSION_MAJOR * 10000 +

aten/src/ATen/cuda/detail/CUDAHooks.h

@@ -49,8 +49,6 @@ struct CUDAHooks : public at::CUDAHooksInterface {
   bool hasCUDART() const override;
   long versionCUDART() const override;
   long versionCuDNN() const override;
-  long versionRuntimeCuDNN() const override;
-  long versionCuDNNFrontend() const override;
   long versionMIOpen() const override;
   std::string showConfig() const override;
   double batchnormMinEpsilonCuDNN() const override;

aten/src/ATen/detail/CUDAHooksInterface.h

@@ -174,14 +174,6 @@ struct TORCH_API CUDAHooksInterface : AcceleratorHooksInterface {
     TORCH_CHECK(false, "Cannot query cuDNN version without ATen_cuda library. ", CUDA_HELP);
   }
 
-  virtual long versionRuntimeCuDNN() const {
-    TORCH_CHECK(false, "Cannot query cuDNN version without ATen_cuda library. ", CUDA_HELP);
-  }
-
-  virtual long versionCuDNNFrontend() const {
-    TORCH_CHECK(false, "Cannot query cuDNN Frontend version without ATen_cuda library. ", CUDA_HELP);
-  }
-
   virtual long versionMIOpen() const {
     TORCH_CHECK(false, "Cannot query MIOpen version without ATen_cuda library. ", CUDA_HELP);
   }

aten/src/ATen/functorch/BatchedTensorImpl.h

@@ -157,8 +157,6 @@ constexpr DispatchKeySet kKeysToPropagateToWrapper({
   DispatchKey::Negative,
   DispatchKey::Conjugate,
   DispatchKey::XLA,
-  DispatchKey::XPU,
-  DispatchKey::HPU,
   DispatchKey::CUDA,
   DispatchKey::CPU,
   DispatchKey::PrivateUse1,

aten/src/ATen/native/Convolution.cpp

@@ -409,7 +409,7 @@ struct ConvParams {
     if (!detail::getCUDAHooks().compiledWithCuDNN() || !input.is_cuda() || !cudnn_enabled) {
       return false;
     }
-    static long cudnn_version = detail::getCUDAHooks().versionRuntimeCuDNN();
+    static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
     // broken on cuDNN 9.8 - 9.14
     if (cudnn_version >= 90800 && cudnn_version < 91500) {
       if (cudnn_conv_suggest_memory_format(input, weight) == at::MemoryFormat::Contiguous &&
@@ -453,7 +453,7 @@ struct ConvParams {
     }
     // native kernel doesn't support 64-bit non-splittable case
     if (!(canUse32BitIndexMath(input) && canUse32BitIndexMath(weight))) {
-      static long cudnn_version = detail::getCUDAHooks().compiledWithCuDNN() ? detail::getCUDAHooks().versionRuntimeCuDNN() : -1;
+      static long cudnn_version = detail::getCUDAHooks().compiledWithCuDNN() ? detail::getCUDAHooks().versionCuDNN() : -1;
       // TODO(eqy): remove this once cuDNN fixes 64-bit depthwise support, first broken in 9.11x
       if (cudnn_conv_suggest_memory_format(input, weight) != at::MemoryFormat::Contiguous) {
         if (cudnn_version < 0 || cudnn_version > 91000) {

aten/src/ATen/native/cuda/CUDALoops.cuh

@@ -884,69 +884,6 @@ struct type_specialized_kernel_launcher {
   }
 };
 
-template <int arg_index>
-struct type_specialized_broadcast_kernel_launcher {
-  template <
-      typename func_t,
-      typename array_t,
-      typename dtypes_t,
-      typename calc_t>
-  static void apply(
-      int64_t numel,
-      func_t f,
-      array_t data,
-      dtypes_t dtypes,
-      calc_t offset_calc) {
-        using traits = function_traits<func_t>;
-        using ret_t = typename traits::result_type;
-        using arg0_t = typename traits::template arg<0>::type;
-        using arg1_t = typename traits::template arg<1>::type;
-        if (dtypes[0] == rt_binary_specializations[arg_index][0] &&
-          dtypes[1] == rt_binary_specializations[arg_index][1] &&
-          dtypes[2] == rt_binary_specializations[arg_index][2]) {
-            using ret_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][0]>;
-            using arg0_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][1]>;
-            using arg1_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][2]>;
-            constexpr int grp_sz = 128;
-            launch_legacy_kernel_manual_unroll<grp_sz, 4>(numel, [=] GPU_LAMBDA(int idx, bool unrl) {
-              if (unrl) {
-                auto offsets0 = offset_calc.get(idx);
-                auto offsets1 = offset_calc.get(idx + grp_sz);
-                auto offsets2 = offset_calc.get(idx + grp_sz * 2);
-                auto offsets3 = offset_calc.get(idx + grp_sz * 3);
-                void* out0 = data[0] + offsets0[0];
-                void* out1 = data[0] + offsets1[0];
-                void* out2 = data[0] + offsets2[0];
-                void* out3 = data[0] + offsets3[0];
-                auto u = c10::load<arg0_cpp_t>(data[1] + offsets0[1]);
-                auto v = c10::load<arg1_cpp_t>(data[2] + offsets0[2]);
-                ret_t result0 = f(c10::convert<arg0_t>(u), c10::convert<arg1_t>(v));
-                auto u1 = c10::load<arg0_cpp_t>(data[1] + offsets1[1]);
-                auto v1 = c10::load<arg1_cpp_t>(data[2]+ offsets1[2]);
-                ret_t result1 = f(c10::convert<arg0_t>(u1), c10::convert<arg1_t>(v1));
-                auto u2 = c10::load<arg0_cpp_t>(data[1] + offsets2[1]);
-                auto v2 = c10::load<arg1_cpp_t>(data[2] + offsets2[2]);
-                ret_t result2 = f(c10::convert<arg0_t>(u2), c10::convert<arg1_t>(v2));
-                auto u3 = c10::load<arg0_cpp_t>(data[1] + offsets3[1]);
-                auto v3 = c10::load<arg1_cpp_t>(data[2] + offsets3[2]);
-                ret_t result3 = f(c10::convert<arg0_t>(u3), c10::convert<arg1_t>(v3));
-                *(ret_cpp_t*)out0 = c10::convert<ret_cpp_t>(result0);
-                *(ret_cpp_t*)out1 = c10::convert<ret_cpp_t>(result1);
-                *(ret_cpp_t*)out2 = c10::convert<ret_cpp_t>(result2);
-                *(ret_cpp_t*)out3 = c10::convert<ret_cpp_t>(result3);
-              } else {
-                auto offsets = offset_calc.get(idx);
-                void* out = data[0] + offsets[0];
-                auto u = c10::load<arg0_cpp_t>(data[1] + offsets[1]);
-                auto v = c10::load<arg1_cpp_t>(data[2] + offsets[2]);
-                ret_t result = f(c10::convert<arg0_t>(u), c10::convert<arg1_t>(v));
-                *(ret_cpp_t*)out = c10::convert<ret_cpp_t>(result);
-              }
-            });
-        }
-      }
-};
-
 } // namespace
 #endif
 
@@ -1065,32 +1002,6 @@ void gpu_kernel_impl(TensorIteratorBase& iter, const func_t& f) {
     }
     auto offset_calc = ::make_offset_calculator<traits::arity + 1>(iter);
 #ifdef USE_ROCM
-    if (check_binary_rt_types_for_specialization(iter)) {
-      // constexpr to reduce the amount of kernels generated for
-      // broadcast elementwise with mexed dtypes and limit which functors are actually
-      // applied to the load and store at compile time.
-      using func_tuple = typename traits::ArgsTuple;
-      if constexpr (
-        std::is_same_v<float, arg0_t> && traits::arity == 2 &&
-        check_binary_functor_types_for_specialization<
-          func_tuple,
-          float,
-          float,
-          traits::arity,
-          /*arg_num=*/0>::check()) {
-            memory::detail::static_unroll<
-              type_specialized_broadcast_kernel_launcher,
-              rt_binary_specializations.size()>::with_args(
-                numel,
-                f,
-                data,
-                dtypes,
-                offset_calc
-            );
-            return;
-      }
-    }
-
     constexpr int grp_sz = 128;
     launch_legacy_kernel_manual_unroll<grp_sz, 4>(numel, [=] GPU_LAMBDA(int idx, bool unrl) {
       if (unrl) {

aten/src/ATen/native/mkldnn/xpu/detail/QConv.cpp

@@ -133,7 +133,7 @@ at::Tensor quantized_convolution(
   // supported in conv.
   mask_weight = weight_zero_points.numel() > 1 ? 1 : 0;
   if (groups > 1 && weight_zero_points.numel() > 1)
-    mask_weight = (1 << 0) | (1 << 1); // 2^0 (group) | 2^1 (output channel)
+    mask_weight = (2 ^ 0) | (2 ^ 1); // 2^0 (group) | 2^1 (output channel)
   dnnl::primitive_attr pattr;
 
   bool src_need_zp = (act_zero_point != 0);

aten/src/ATen/native/mps/operations/Blas.mm

@@ -141,9 +141,6 @@ static Tensor& addmv_out_mps_impl(const Tensor& self,
   };
 
   MPSStream* stream = at::mps::getCurrentMPSStream();
-  if (result.numel() == 0) {
-    return result;
-  }
   Tensor matMulVec = at::mm(mat, vec.unsqueeze(1)).squeeze(1);
 
   @autoreleasepool {

aten/src/ATen/native/native_functions.yaml

@@ -2803,7 +2803,7 @@
 - func: floor_divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   dispatch:
-    CPU, CUDA, MPS, MTIA: floor_divide_out
+    CPU, CUDA, MPS: floor_divide_out
     SparseCPU, SparseCUDA, SparseMPS: floor_divide_out_sparse_zerodim
 
 - func: floor_divide.Scalar(Tensor self, Scalar other) -> Tensor
@@ -4292,7 +4292,6 @@
   dispatch:
     SparseCPU: sparse_sparse_matmul_cpu
     SparseCUDA: sparse_sparse_matmul_cuda
-    SparseMPS: sparse_sparse_matmul_mps
   autogen: _sparse_sparse_matmul.out
 
 - func: mode(Tensor self, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices)
@@ -4384,7 +4383,7 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: mv
-    SparseCPU, SparseCUDA, SparseMPS: mv_sparse
+    SparseCPU, SparseCUDA: mv_sparse
 
 - func: mv.out(Tensor self, Tensor vec, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
@@ -9833,7 +9832,7 @@
   structured_delegate: erfinv.out
   variants: method, function
   dispatch:
-    SparseCPU, SparseCUDA, SparseMPS: erfinv_sparse
+    SparseCPU, SparseCUDA: erfinv_sparse
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: erfinv_sparse_csr
   tags: pointwise
 
@@ -9842,7 +9841,7 @@
   structured_delegate: erfinv.out
   variants: method
   dispatch:
-    SparseCPU, SparseCUDA, SparseMPS: erfinv_sparse_
+    SparseCPU, SparseCUDA: erfinv_sparse_
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: erfinv_sparse_csr_
   tags: pointwise
 
@@ -9852,7 +9851,7 @@
   structured_inherits: TensorIteratorBase
   dispatch:
     CPU, CUDA, MPS: erfinv_out
-    SparseCPU, SparseCUDA, SparseMPS: erfinv_sparse_out
+    SparseCPU, SparseCUDA: erfinv_sparse_out
     SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: erfinv_sparse_csr_out
   tags: pointwise
 

aten/src/ATen/native/sparse/mps/SparseMPSTensorMath.mm

@@ -10,10 +10,6 @@
 #include <ATen/NativeFunctions.h>
 #else
 #include <ATen/ops/_coalesce_native.h>
-#include <ATen/ops/repeat_interleave_native.h>
-#include <ATen/ops/cumsum.h>
-#include <ATen/ops/_sparse_sparse_matmul_native.h>
-#include <ATen/ops/_sparse_coo_tensor_unsafe.h>
 #include <ATen/ops/_sparse_coo_tensor_unsafe_native.h>
 #include <ATen/ops/cat.h>
 #include <ATen/ops/add_native.h>
@@ -892,114 +888,5 @@ static void sparse_mask_intersection_out_mps_kernel(
       /*coalesce_mask=*/false);
 }
 
-Tensor sparse_sparse_matmul_mps(const Tensor& mat1_, const Tensor& mat2_) {
-  TORCH_CHECK(mat1_.is_sparse() && mat2_.is_sparse(),
-              "sparse_sparse_matmul_mps: both inputs must be sparse COO tensors");
-  TORCH_CHECK(mat1_.is_mps() && mat2_.is_mps(),
-              "sparse_sparse_matmul_mps: both inputs must be on MPS device");
-  TORCH_CHECK(mat1_.dim() == 2 && mat2_.dim() == 2,
-              "sparse_sparse_matmul_mps: both inputs must be 2D matrices");
-  TORCH_CHECK(mat1_.dense_dim() == 0 && mat2_.dense_dim() == 0,
-              "sparse_sparse_matmul_mps: only scalar values supported (dense_dim == 0)");
-  TORCH_CHECK(mat1_.size(1) == mat2_.size(0),
-              "mat1 and mat2 shapes cannot be multiplied (", mat1_.size(0), "x", mat1_.size(1), " and ", mat2_.size(0), "x", mat2_.size(1), ")");
-  TORCH_CHECK(mat1_.scalar_type() == mat2_.scalar_type(),
-              "sparse_sparse_matmul_mps: mat1 dtype ", mat1_.scalar_type(),
-              " does not match mat2 dtype ", mat2_.scalar_type());
-
-  const auto device = mat1_.device();
-
-  auto A = mat1_.coalesce();
-  auto B = mat2_.coalesce();
-
-  const auto I = A.size(0);
-  const auto K = A.size(1);
-  const auto N = B.size(1);
-
-  const auto nnzA = A._nnz();
-  const auto nnzB = B._nnz();
-
-  // Early empty result, return an empty, coalesced tensor
-  if (I == 0 || N == 0 || K == 0 || nnzA == 0 || nnzB == 0) {
-    auto empty_idx = at::empty({2, 0}, at::device(device).dtype(at::kLong));
-    auto empty_val = at::empty({0}, at::device(device).dtype(mat1_.scalar_type()));
-    auto out = _sparse_coo_tensor_unsafe(empty_idx, empty_val, {I, N}, mat1_.options());
-    out._coalesced_(true);
-    return out;
-  }
-
-  const auto computeDtype = at::result_type(mat1_, mat2_);
-
-  auto A_idx = A._indices().contiguous();
-  auto A_val = A._values().to(computeDtype).contiguous();
-  auto A_i = A_idx.select(0, 0).contiguous();
-  auto A_k = A_idx.select(0, 1).contiguous();
-
-  auto B_idx = B._indices().contiguous();
-  auto B_val = B._values().to(computeDtype).contiguous();
-  auto B_k = B_idx.select(0, 0).contiguous();
-  auto B_j = B_idx.select(0, 1).contiguous();
-
-  // csr-style row pointers for B by k (the shared dimension)
-  Tensor row_ptr_B;
-  {
-    auto batch_ptr = at::tensor({0LL, nnzB}, at::device(device).dtype(at::kLong));
-    row_ptr_B = at::empty({K + 1}, at::device(device).dtype(at::kLong));
-    build_row_ptr_per_batch_mps(B_k, batch_ptr, /*B=*/1, /*I=*/K, row_ptr_B);
-  }
-
-  auto row_ptr_B_lo = row_ptr_B.narrow(0, 0, K);
-  auto row_ptr_B_hi = row_ptr_B.narrow(0, 1, K);
-  auto deg_B = row_ptr_B_hi.sub(row_ptr_B_lo);
-
-  auto counts = deg_B.index_select(0, A_k);
-
-  const int64_t P = counts.sum().item<int64_t>();
-  if (P == 0) {
-    auto empty_idx = at::empty({2, 0}, at::device(device).dtype(at::kLong));
-    auto empty_val = at::empty({0}, at::device(device).dtype(mat1_.scalar_type()));
-    auto out = _sparse_coo_tensor_unsafe(empty_idx, empty_val, {I, N}, mat1_.options());
-    out._coalesced_(true);
-    return out;
-  }
-
-  auto group_ids = repeat_interleave_mps(counts);
-
-  // exclusive cumsum of counts
-  auto offsets = cumsum(counts, /*dim=*/0).sub(counts);
-  auto offsets_gather = offsets.index_select(0, group_ids);
-  auto within = at::arange(P, at::device(device).dtype(at::kLong)).sub(offsets_gather);
-
-  // Map each output element to its source B row and position
-  auto k_per_out = A_k.index_select(0, group_ids);
-  auto start_in_B = row_ptr_B.index_select(0, k_per_out);
-  auto seg_index = start_in_B.add(within);
-
-  // Assemble candidate coo pairs and values
-  auto i_out = A_i.index_select(0, group_ids).contiguous();
-  auto j_out = B_j.index_select(0, seg_index).contiguous();
-  auto vA_out = A_val.index_select(0, group_ids).contiguous();
-  auto vB_out = B_val.index_select(0, seg_index).contiguous();
-  auto v_out = vA_out.mul(vB_out);
-
-  // build (2, P) indices
-  auto out_indices = at::empty({2, P}, at::device(device).dtype(at::kLong)).contiguous();
-  out_indices.select(0, 0).copy_(i_out);
-  out_indices.select(0, 1).copy_(j_out);
-
-  auto result = _sparse_coo_tensor_unsafe(
-      out_indices, v_out, {I, N}, mat1_.options().dtype(computeDtype));
-
-  result = result.coalesce();
-
-  if (result.scalar_type() != mat1_.scalar_type()) {
-    auto cast_vals = result._values().to(mat1_.scalar_type());
-    auto out = _sparse_coo_tensor_unsafe(result._indices(), cast_vals, {I, N}, mat1_.options());
-    out._coalesced_(true);
-    return out;
-  }
-  return result;
-}
-
 REGISTER_MPS_DISPATCH(sparse_mask_intersection_out_stub, &sparse_mask_intersection_out_mps_kernel);
 } // namespace at::native

aten/src/ATen/native/transformers/cuda/sdp_utils.cpp

@@ -478,7 +478,7 @@ bool check_cudnn_tensor_shapes(sdp_params const& params, bool debug) {
   const auto s_k = params.key.sym_size(2);
   const auto d_qk = params.query.sym_size(3);
   const auto d_v = params.value.sym_size(3);
-  long cudnn_version = at::detail::getCUDAHooks().versionRuntimeCuDNN();
+  long cudnn_version = at::detail::getCUDAHooks().versionCuDNN();
   if (cudnn_version < 8903) {
     if (debug) {
       TORCH_WARN("SDPA fprop requires cudnn 8.9.3 or higher");
@@ -709,7 +709,7 @@ bool can_use_cudnn_attention(const sdp_params& params, bool debug) {
   return false;
 #endif
 #if defined(CUDNN_VERSION)
-  static auto cudnn_version = at::detail::getCUDAHooks().versionRuntimeCuDNN();
+  static auto cudnn_version = cudnnGetVersion();
   if (params.dropout > 0.0 && cudnn_version > 91100 && cudnn_version < 91400) {
     if (debug) {
       TORCH_WARN(CUDNN_VERSION, " cuDNN version does not support droppout in SDPA (9.11 - 9.13).");

benchmarks/sparse/spmm.py

@@ -52,18 +52,19 @@ def test_sparse_coo_and_csr(m, n, k, nnz, test_count):
         start.record()
         coo.matmul(mat)
         stop.record()
+
         times.append(start.elapsed_time(stop))
 
-    coo_mean_time = sum(times) / len(times)
+        coo_mean_time = sum(times) / len(times)
 
-    times = []
-    for _ in range(test_count):
-        start.record()
-        csr.matmul(mat)
-        stop.record()
-        times.append(start.elapsed_time(stop))
+        times = []
+        for _ in range(test_count):
+            start.record()
+            csr.matmul(mat)
+            stop.record()
+            times.append(start.elapsed_time(stop))
 
-    csr_mean_time = sum(times) / len(times)
+            csr_mean_time = sum(times) / len(times)
 
     return coo_mean_time, csr_mean_time
 

c10/core/AutogradState.h

@@ -1,8 +1,6 @@
 #pragma once
 
-#include <c10/core/SafePyObject.h>
 #include <c10/macros/Export.h>
-#include <optional>
 
 namespace c10 {
 
@@ -17,8 +15,7 @@ struct C10_API AutogradState {
       bool inference_mode,
       bool fw_grad_mode,
       bool multithreading_enabled)
-      : graph_exec_group_(std::nullopt),
-        grad_mode_(grad_mode),
+      : grad_mode_(grad_mode),
         inference_mode_(inference_mode),
         fw_grad_mode_(fw_grad_mode),
         multithreading_enabled_(multithreading_enabled),
@@ -44,10 +41,6 @@ struct C10_API AutogradState {
     view_replay_enabled_ = view_replay_enabled;
   }
 
-  void set_graph_exec_group(std::optional<SafePyObject> group) {
-    graph_exec_group_ = std::move(group);
-  }
-
   bool get_grad_mode() const {
     return grad_mode_;
   }
@@ -68,12 +61,7 @@ struct C10_API AutogradState {
     return view_replay_enabled_;
   }
 
-  const std::optional<SafePyObject>& get_graph_exec_group() const {
-    return graph_exec_group_;
-  }
-
  private:
-  std::optional<SafePyObject> graph_exec_group_;
   bool grad_mode_ : 1;
   bool inference_mode_ : 1;
   bool fw_grad_mode_ : 1;

c10/core/CachingDeviceAllocator.h

@@ -96,10 +96,6 @@ struct C10_API DeviceAllocator : public c10::Allocator {
 
   // Resets peak memory usage statistics for the specified device
   virtual void resetPeakStats(c10::DeviceIndex device) = 0;
-
-  // Return the free memory size and total memory size in bytes for the
-  // specified device.
-  virtual std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) = 0;
 };
 
 // This function is used to get the DeviceAllocator for a specific device type

c10/cuda/CUDACachingAllocator.h

@@ -345,13 +345,6 @@ class CUDAAllocator : public DeviceAllocator {
       c10::DeviceIndex device,
       std::shared_ptr<AllocatorState> pps) = 0;
   virtual std::string name() = 0;
-  std::pair<size_t, size_t> getMemoryInfo(c10::DeviceIndex device) override {
-    c10::DeviceGuard device_guard({at::kCUDA, device});
-    size_t free = 0;
-    size_t total = 0;
-    C10_CUDA_CHECK(cudaMemGetInfo(&free, &total));
-    return {free, total};
-  }
 };
 
 // Allocator object, statically initialized

c10/test/build.bzl

@@ -66,15 +66,6 @@ def define_targets(rules):
         ],
     )
 
-    rules.cc_test(
-        name = "util/nofatal_test",
-        srcs = ["util/nofatal_test.cpp"],
-        deps = [
-            "//c10/util:base",
-            "@com_google_googletest//:gtest_main",
-        ],
-    )
-
     rules.cc_test(
         name = "util/ssize_test",
         srcs = ["util/ssize_test.cpp"],

c10/test/util/nofatal_test.cpp

@@ -1,53 +0,0 @@
-#include <gtest/gtest.h>
-
-#include <c10/util/Exception.h>
-#include <c10/util/Logging.h>
-
-namespace {
-template <typename T>
-inline void expectThrowsEq(T&& fn, const char* expected_msg) {
-  try {
-    std::forward<T>(fn)();
-  } catch (const c10::Error& e) {
-    EXPECT_TRUE(
-        std::string(e.what_without_backtrace()).find(expected_msg) !=
-        std::string::npos);
-    return;
-  }
-  ADD_FAILURE() << "Expected to throw exception with message \"" << expected_msg
-                << "\" but didn't throw";
-}
-} // namespace
-
-TEST(NofatalTest, TorchCheckComparisons) {
-  // quick make sure that no-op works as expected
-  TORCH_CHECK_EQ(1, 1) << "i am a silly message " << 1;
-  expectThrowsEq(
-      []() { TORCH_CHECK_EQ(1, 2) << "i am a silly message " << 1; },
-      "Check failed: 1 == 2 (1 vs. 2). i am a silly message 1");
-  expectThrowsEq(
-      []() { TORCH_CHECK_NE(2, 2); }, "Check failed: 2 != 2 (2 vs. 2).");
-  expectThrowsEq(
-      []() { TORCH_CHECK_LT(2, 2); }, "Check failed: 2 < 2 (2 vs. 2).");
-  expectThrowsEq(
-      []() { TORCH_CHECK_LE(3, 2); }, "Check failed: 3 <= 2 (3 vs. 2).");
-  expectThrowsEq(
-      []() { TORCH_CHECK_GT(2, 2); }, "Check failed: 2 > 2 (2 vs. 2).");
-  expectThrowsEq(
-      []() { TORCH_CHECK_GE(2, 3); }, "Check failed: 2 >= 3 (2 vs. 3).");
-  expectThrowsEq(
-      []() {
-        void* p = nullptr;
-        TORCH_CHECK_NOTNULL(p);
-      },
-      "Check failed: 'p' must be non NULL.");
-
-#if GTEST_HAS_DEATH_TEST
-#ifndef NDEBUG
-  // if dbg build, DCHECK should result in deth
-  EXPECT_DEATH(TORCH_DCHECK_EQ(1, 2), "Check failed");
-#else
-  TORCH_DCHECK_EQ(1, 2); // no-op
-#endif
-#endif // GTEST_HAS_DEATH_TEST
-}

c10/util/Exception.h

@@ -702,98 +702,6 @@ namespace c10::detail {
 #define TORCH_CHECK_ARG(cond, argN, ...) \
   TORCH_CHECK(cond, "invalid argument ", argN, ": ", __VA_ARGS__)
 
-#ifndef FATAL_IF
-#ifdef C10_USE_GLOG
-#define FATAL_IF(condition)                                              \
-  condition ? (void)0                                                    \
-            : ::c10::LoggerVoidify() &                                   \
-          ::c10::MessageLogger(__FILE__, __LINE__, ::google::GLOG_FATAL) \
-              .stream()
-#else
-#define FATAL_IF(condition)            \
-  condition ? (void)0                  \
-            : ::c10::LoggerVoidify() & \
-          ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_FATAL).stream()
-#endif
-#endif
-
-#ifndef NON_FATAL_IF
-#ifdef C10_USE_GLOG
-#define NON_FATAL_IF(condition)                                \
-  condition ? (void)0                                          \
-            : ::c10::LoggerVoidify() &                         \
-          ::c10::MessageLogger(                                \
-              __FILE__, __LINE__, ::google::GLOG_FATAL, false) \
-              .stream()
-#else
-#define NON_FATAL_IF(condition)                                              \
-  condition ? (void)0                                                        \
-            : ::c10::LoggerVoidify() &                                       \
-          ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_FATAL, false) \
-              .stream()
-#endif
-#endif
-
-// Binary comparison check macros
-#define TORCH_CHECK_OP(val1, val2, op)                                      \
-  NON_FATAL_IF(((val1)op(val2)))                                            \
-      << "Check failed: " #val1 " " #op " " #val2 " (" << (val1) << " vs. " \
-      << (val2) << "). "
-
-#define TORCH_DCHECK_OP(val1, val2, op)                                       \
-  FATAL_IF(((val1)op(val2))) << "Check failed: " #val1 " " #op " " #val2 " (" \
-                             << (val1) << " vs. " << (val2) << "). "
-
-#define TORCH_CHECK_EQ(val1, val2) TORCH_CHECK_OP(val1, val2, ==)
-#define TORCH_CHECK_NE(val1, val2) TORCH_CHECK_OP(val1, val2, !=)
-#define TORCH_CHECK_LE(val1, val2) TORCH_CHECK_OP(val1, val2, <=)
-#define TORCH_CHECK_LT(val1, val2) TORCH_CHECK_OP(val1, val2, <)
-#define TORCH_CHECK_GE(val1, val2) TORCH_CHECK_OP(val1, val2, >=)
-#define TORCH_CHECK_GT(val1, val2) TORCH_CHECK_OP(val1, val2, >)
-
-// Debug versions of TORCH_CHECK_OP macros
-#ifndef NDEBUG
-#define TORCH_DCHECK_EQ(val1, val2) TORCH_DCHECK_OP(val1, val2, ==)
-#define TORCH_DCHECK_NE(val1, val2) TORCH_DCHECK_OP(val1, val2, !=)
-#define TORCH_DCHECK_LE(val1, val2) TORCH_DCHECK_OP(val1, val2, <=)
-#define TORCH_DCHECK_LT(val1, val2) TORCH_DCHECK_OP(val1, val2, <)
-#define TORCH_DCHECK_GE(val1, val2) TORCH_DCHECK_OP(val1, val2, >=)
-#define TORCH_DCHECK_GT(val1, val2) TORCH_DCHECK_OP(val1, val2, >)
-#else // !NDEBUG
-// Optimized versions - generate no code
-#define TORCH_DCHECK_EQ(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, ==)
-#define TORCH_DCHECK_NE(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, !=)
-#define TORCH_DCHECK_LE(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, <=)
-#define TORCH_DCHECK_LT(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, <)
-#define TORCH_DCHECK_GE(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, >=)
-#define TORCH_DCHECK_GT(val1, val2) \
-  while (false)                     \
-  TORCH_DCHECK_OP(val1, val2, >)
-#endif // NDEBUG
-
-// Null pointer check macro
-#define TORCH_CHECK_NOTNULL(val) \
-  ::c10::CheckNotNull(__FILE__, __LINE__, #val, (val), false)
-
-#ifndef NDEBUG
-#define TORCH_DCHECK_NOTNULL(val) \
-  ::c10::CheckNotNull(__FILE__, __LINE__, #val, (val), true)
-#else // !NDEBUG
-#define TORCH_DCHECK_NOTNULL(val) \
-  while (false)                   \
-  TORCH_CHECK_NOTNULL(val)
-#endif // NDEBUG
-
 // ----------------------------------------------------------------------------
 // Deprecated macros
 // ----------------------------------------------------------------------------

c10/util/Logging.cpp

@@ -291,32 +291,6 @@ namespace c10 {
 using fLB::FLAGS_logtostderr;
 using fLI::FLAGS_minloglevel;
 using fLI::FLAGS_v;
-
-MessageLogger::MessageLogger(
-    const char* file,
-    int line,
-    int severity,
-    bool exit_on_fatal)
-    : stream_(), severity_(severity), exit_on_fatal_(exit_on_fatal) {}
-
-MessageLogger::~MessageLogger() noexcept(false) {
-  if (severity_ == ::google::GLOG_FATAL) {
-    DealWithFatal();
-  }
-}
-
-std::stringstream& MessageLogger::stream() {
-  return stream_;
-}
-
-void MessageLogger::DealWithFatal() {
-  if (exit_on_fatal_) {
-    LOG(FATAL) << stream_.str();
-  } else {
-    throw c10::Error(stream_.str(), nullptr, nullptr);
-  }
-}
-
 } // namespace c10
 
 C10_DEFINE_int(
@@ -438,16 +412,17 @@ void ShowLogInfoToStderr() {
   FLAGS_caffe2_log_level = GLOG_INFO;
 }
 
-MessageLogger::MessageLogger(
-    const char* file,
-    int line,
-    int severity,
-    bool exit_on_fatal)
-    : severity_(severity), exit_on_fatal_(exit_on_fatal) {
+MessageLogger::MessageLogger(const char* file, int line, int severity)
+    : severity_(severity) {
   if (severity_ < FLAGS_caffe2_log_level) {
     // Nothing needs to be logged.
     return;
   }
+#ifdef ANDROID
+  tag_ = "native";
+#else // !ANDROID
+  tag_ = "";
+#endif // ANDROID
 
   time_t rawtime = 0;
   time(&rawtime);
@@ -483,7 +458,7 @@ MessageLogger::MessageLogger(
 }
 
 // Output the contents of the stream to the proper channel on destruction.
-MessageLogger::~MessageLogger() noexcept(false) {
+MessageLogger::~MessageLogger() {
   if (severity_ < FLAGS_caffe2_log_level) {
     // Nothing needs to be logged.
     return;
@@ -523,18 +498,6 @@ MessageLogger::~MessageLogger() noexcept(false) {
   }
 }
 
-std::stringstream& MessageLogger::stream() {
-  return stream_;
-}
-
-void MessageLogger::DealWithFatal() {
-  if (exit_on_fatal_) {
-    abort();
-  } else {
-    throw c10::Error(stream_.str(), nullptr, nullptr);
-  }
-}
-
 } // namespace c10
 
 #endif // !C10_USE_GLOG

c10/util/logging_common.h

@@ -1,74 +0,0 @@
-#ifndef C10_UTIL_LOGGING_COMMON_H_
-#define C10_UTIL_LOGGING_COMMON_H_
-
-#include <c10/macros/Export.h>
-#include <sstream>
-
-namespace c10 {
-
-// MessageLogger that throws exceptions instead of aborting (glog version)
-// or logs and may abort (non-glog version).
-class C10_API MessageLogger {
- public:
-  MessageLogger(
-      const char* file,
-      int line,
-      int severity,
-      bool exit_on_fatal = true);
-  ~MessageLogger() noexcept(false);
-
-  // Return the stream associated with the logger object.
-  std::stringstream& stream();
-
- private:
-  // When there is a fatal log, and fatal == true, we abort
-  // otherwise, we throw.
-  void DealWithFatal();
-
-#if defined(ANDROID) && !defined(C10_USE_GLOG)
-  const char* tag_{"native"};
-#endif
-  std::stringstream stream_;
-  int severity_;
-  bool exit_on_fatal_;
-};
-
-// This class is used to explicitly ignore values in the conditional
-// logging macros. This avoids compiler warnings like "value computed
-// is not used" and "statement has no effect".
-class C10_API LoggerVoidify {
- public:
-  LoggerVoidify() = default;
-  // This has to be an operator with a precedence lower than << but
-  // higher than ?:
-  void operator&(const std::ostream& s [[maybe_unused]]) {}
-};
-
-// Forward declarations for CheckNotNull functions
-template <typename T>
-T& CheckNotNullCommon(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal = true);
-
-template <typename T>
-T* CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T* t,
-    bool fatal = true);
-
-template <typename T>
-T& CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal = true);
-
-} // namespace c10
-
-#endif // C10_UTIL_LOGGING_COMMON_H_

c10/util/logging_is_google_glog.h

@@ -47,53 +47,57 @@ INSTANTIATE_FOR_CONTAINER(set)
 
 #endif
 
-#include <c10/util/logging_common.h>
 #include <glog/logging.h>
 
-namespace c10 {
+// Additional macros on top of glog
+#define TORCH_CHECK_EQ(val1, val2) CHECK_EQ(val1, val2)
+#define TORCH_CHECK_NE(val1, val2) CHECK_NE(val1, val2)
+#define TORCH_CHECK_LE(val1, val2) CHECK_LE(val1, val2)
+#define TORCH_CHECK_LT(val1, val2) CHECK_LT(val1, val2)
+#define TORCH_CHECK_GE(val1, val2) CHECK_GE(val1, val2)
+#define TORCH_CHECK_GT(val1, val2) CHECK_GT(val1, val2)
 
-[[noreturn]] void ThrowEnforceNotMet(
-    const char* file,
-    const int line,
-    const char* condition,
-    const std::string& msg,
-    const void* caller);
+#ifndef NDEBUG
+#define TORCH_DCHECK_EQ(val1, val2) DCHECK_EQ(val1, val2)
+#define TORCH_DCHECK_NE(val1, val2) DCHECK_NE(val1, val2)
+#define TORCH_DCHECK_LE(val1, val2) DCHECK_LE(val1, val2)
+#define TORCH_DCHECK_LT(val1, val2) DCHECK_LT(val1, val2)
+#define TORCH_DCHECK_GE(val1, val2) DCHECK_GE(val1, val2)
+#define TORCH_DCHECK_GT(val1, val2) DCHECK_GT(val1, val2)
+#else // !NDEBUG
+// These versions generate no code in optimized mode.
+#define TORCH_DCHECK_EQ(val1, val2) \
+  while (false)                     \
+  DCHECK_EQ(val1, val2)
+#define TORCH_DCHECK_NE(val1, val2) \
+  while (false)                     \
+  DCHECK_NE(val1, val2)
+#define TORCH_DCHECK_LE(val1, val2) \
+  while (false)                     \
+  DCHECK_LE(val1, val2)
+#define TORCH_DCHECK_LT(val1, val2) \
+  while (false)                     \
+  DCHECK_LT(val1, val2)
+#define TORCH_DCHECK_GE(val1, val2) \
+  while (false)                     \
+  DCHECK_GE(val1, val2)
+#define TORCH_DCHECK_GT(val1, val2) \
+  while (false)                     \
+  DCHECK_GT(val1, val2)
+#endif // NDEBUG
 
-template <typename T>
-T& CheckNotNullCommon(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal) {
-  if (t == nullptr) {
-    MessageLogger(file, line, ::google::GLOG_FATAL, fatal).stream()
-        << "Check failed: '" << names << "' must be non NULL. ";
-  }
-  return t;
-}
+// Check that a pointer is not null.
+#define TORCH_CHECK_NOTNULL(val) CHECK_NOTNULL(val)
 
-template <typename T>
-T* CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T* t,
-    bool fatal) {
-  return CheckNotNullCommon(file, line, names, t, fatal);
-}
-
-template <typename T>
-T& CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal) {
-  return CheckNotNullCommon(file, line, names, t, fatal);
-}
-
-} // namespace c10
+#ifndef NDEBUG
+// Debug only version of TORCH_CHECK_NOTNULL
+#define TORCH_DCHECK_NOTNULL(val) DCHECK_NOTNULL(val)
+#else // !NDEBUG
+// Optimized version - generates no code.
+#define TORCH_DCHECK_NOTNULL(val) \
+  while (false)                   \
+  DCHECK_NOTNULL(val)
+#endif // NDEBUG
 
 // Log with source location information override (to be used in generic
 // warning/error handlers implemented as functions, not macros)

c10/util/logging_is_not_google_glog.h

@@ -13,7 +13,6 @@
 #include <vector>
 
 #include <c10/util/Flags.h>
-#include <c10/util/logging_common.h>
 
 const char CAFFE2_SEVERITY_PREFIX[] = "FEWIV";
 
@@ -25,40 +24,61 @@ const int GLOG_ERROR = 2;
 const int GLOG_WARNING = 1;
 const int GLOG_INFO = 0;
 
+class C10_API MessageLogger {
+ public:
+  MessageLogger(const char* file, int line, int severity);
+  ~MessageLogger();
+  // Return the stream associated with the logger object.
+  std::stringstream& stream() {
+    return stream_;
+  }
+
+ private:
+  // When there is a fatal log, we simply abort.
+  void DealWithFatal() {
+    abort();
+  }
+
+  const char* tag_;
+  std::stringstream stream_;
+  int severity_;
+};
+
+// This class is used to explicitly ignore values in the conditional
+// logging macros.  This avoids compiler warnings like "value computed
+// is not used" and "statement has no effect".
+class C10_API LoggerVoidify {
+ public:
+  LoggerVoidify() = default;
+  // This has to be an operator with a precedence lower than << but
+  // higher than ?:
+  void operator&(const std::ostream& s [[maybe_unused]]) {}
+};
+
+// Log a message and terminate.
+template <class T>
+void LogMessageFatal(const char* file, int line, const T& message) {
+  MessageLogger(file, line, GLOG_FATAL).stream() << message;
+}
+
 // Helpers for TORCH_CHECK_NOTNULL(). Two are necessary to support both raw
 // pointers and smart pointers.
 template <typename T>
-T& CheckNotNullCommon(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal) {
+T& CheckNotNullCommon(const char* file, int line, const char* names, T& t) {
   if (t == nullptr) {
-    MessageLogger(file, line, GLOG_FATAL, fatal).stream()
-        << "Check failed: '" << names << "' must be non NULL. ";
+    LogMessageFatal(file, line, std::string(names));
   }
   return t;
 }
 
 template <typename T>
-T* CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T* t,
-    bool fatal) {
-  return CheckNotNullCommon(file, line, names, t, fatal);
+T* CheckNotNull(const char* file, int line, const char* names, T* t) {
+  return CheckNotNullCommon(file, line, names, t);
 }
 
 template <typename T>
-T& CheckNotNull(
-    const char* file,
-    int line,
-    const char* names,
-    T& t,
-    bool fatal) {
-  return CheckNotNullCommon(file, line, names, t, fatal);
+T& CheckNotNull(const char* file, int line, const char* names, T& t) {
+  return CheckNotNullCommon(file, line, names, t);
 }
 } // namespace c10
 
@@ -116,6 +136,65 @@ static_assert(
           ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_##n).stream()
 #endif // NDEBUG
 
+#define TORCH_CHECK_OP(val1, val2, op)                                        \
+  FATAL_IF(((val1)op(val2))) << "Check failed: " #val1 " " #op " " #val2 " (" \
+                             << (val1) << " vs. " << (val2) << ") "
+
+// TORCH_CHECK_OP macro definitions
+#define TORCH_CHECK_EQ(val1, val2) TORCH_CHECK_OP(val1, val2, ==)
+#define TORCH_CHECK_NE(val1, val2) TORCH_CHECK_OP(val1, val2, !=)
+#define TORCH_CHECK_LE(val1, val2) TORCH_CHECK_OP(val1, val2, <=)
+#define TORCH_CHECK_LT(val1, val2) TORCH_CHECK_OP(val1, val2, <)
+#define TORCH_CHECK_GE(val1, val2) TORCH_CHECK_OP(val1, val2, >=)
+#define TORCH_CHECK_GT(val1, val2) TORCH_CHECK_OP(val1, val2, >)
+
+#ifndef NDEBUG
+// Debug only versions of TORCH_CHECK_OP macros.
+#define TORCH_DCHECK_EQ(val1, val2) TORCH_CHECK_OP(val1, val2, ==)
+#define TORCH_DCHECK_NE(val1, val2) TORCH_CHECK_OP(val1, val2, !=)
+#define TORCH_DCHECK_LE(val1, val2) TORCH_CHECK_OP(val1, val2, <=)
+#define TORCH_DCHECK_LT(val1, val2) TORCH_CHECK_OP(val1, val2, <)
+#define TORCH_DCHECK_GE(val1, val2) TORCH_CHECK_OP(val1, val2, >=)
+#define TORCH_DCHECK_GT(val1, val2) TORCH_CHECK_OP(val1, val2, >)
+#else // !NDEBUG
+// These versions generate no code in optimized mode.
+#define TORCH_DCHECK_EQ(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, ==)
+#define TORCH_DCHECK_NE(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, !=)
+#define TORCH_DCHECK_LE(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, <=)
+#define TORCH_DCHECK_LT(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, <)
+#define TORCH_DCHECK_GE(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, >=)
+#define TORCH_DCHECK_GT(val1, val2) \
+  while (false)                     \
+  TORCH_CHECK_OP(val1, val2, >)
+#endif // NDEBUG
+
+// Check that a pointer is not null.
+#define TORCH_CHECK_NOTNULL(val) \
+  ::c10::CheckNotNull(           \
+      __FILE__, __LINE__, "Check failed: '" #val "' Must be non NULL", (val))
+
+#ifndef NDEBUG
+// Debug only version of TORCH_CHECK_NOTNULL
+#define TORCH_DCHECK_NOTNULL(val) \
+  ::c10::CheckNotNull(            \
+      __FILE__, __LINE__, "Check failed: '" #val "' Must be non NULL", (val))
+#else // !NDEBUG
+// Optimized version - generates no code.
+#define TORCH_DCHECK_NOTNULL(val) \
+  while (false)                   \
+  TORCH_CHECK_NOTNULL(val)
+#endif // NDEBUG
+
 // ---------------------- Support for std objects --------------------------
 // These are adapted from glog to support a limited set of logging capability
 // for STL objects.

c10/xpu/XPUCachingAllocator.cpp

@@ -926,14 +926,15 @@ class DeviceCachingAllocator {
           (release_cached_blocks() && alloc_block(params, true));
     }
     if (!block_found) {
-      const auto& raw_device = c10::xpu::get_raw_device(device);
-      const auto device_total =
-          raw_device.get_info<sycl::info::device::global_mem_size>();
+      c10::xpu::DeviceProp device_prop;
+      c10::xpu::get_device_properties(&device_prop, device);
+      auto device_total = device_prop.global_mem_size;
       // Estimate the available device memory when the SYCL runtime does not
       // support the corresponding aspect (ext_intel_free_memory).
-      size_t device_free = device_total -
+      size_t device_free = device_prop.global_mem_size -
           stats.reserved_bytes[static_cast<size_t>(StatType::AGGREGATE)]
               .current;
+      auto& raw_device = c10::xpu::get_raw_device(device);
       // TODO: Remove the aspect check once the SYCL runtime bug is fixed on
       // affected devices.
       if (raw_device.has(sycl::aspect::ext_intel_free_memory)) {
@@ -1051,37 +1052,21 @@ class DeviceCachingAllocator {
     }
   }
 
-  std::pair<size_t, size_t> getMemoryInfo() {
-    const auto& device = c10::xpu::get_raw_device(device_index);
-    const size_t total = device.get_info<sycl::info::device::global_mem_size>();
-    TORCH_CHECK(
-        device.has(sycl::aspect::ext_intel_free_memory),
-        "The device (",
-        device.get_info<sycl::info::device::name>(),
-        ") doesn't support querying the available free memory. ",
-        "You can file an issue at https://github.com/pytorch/pytorch/issues ",
-        "to help us prioritize its implementation.");
-    const size_t free =
-        device.get_info<sycl::ext::intel::info::device::free_memory>();
-    return {free, total};
-  }
-
   double getMemoryFraction() {
     if (!set_fraction) {
       return 1.0;
     }
 
-    const auto device_total =
-        xpu::get_raw_device(device_index)
-            .get_info<sycl::info::device::global_mem_size>();
+    c10::xpu::DeviceProp device_prop;
+    c10::xpu::get_device_properties(&device_prop, device_index);
     return static_cast<double>(allowed_memory_maximum) /
-        static_cast<double>(device_total);
+        static_cast<double>(device_prop.global_mem_size);
   }
 
   void setMemoryFraction(double fraction) {
-    const auto device_total =
-        xpu::get_raw_device(device_index)
-            .get_info<sycl::info::device::global_mem_size>();
+    c10::xpu::DeviceProp device_prop;
+    c10::xpu::get_device_properties(&device_prop, device_index);
+    auto device_total = device_prop.global_mem_size;
     allowed_memory_maximum = static_cast<size_t>(fraction * device_total);
     set_fraction = true;
   }
@@ -1255,11 +1240,6 @@ class XPUAllocator : public DeviceAllocator {
         c10::xpu::get_raw_device(dev_to_access));
   }
 
-  std::pair<size_t, size_t> getMemoryInfo(DeviceIndex device) override {
-    assertValidDevice(device);
-    return device_allocators[device]->getMemoryInfo();
-  }
-
   double getMemoryFraction(DeviceIndex device) {
     assertValidDevice(device);
     return device_allocators[device]->getMemoryFraction();

caffe2/CMakeLists.txt

@@ -1941,7 +1941,6 @@ if(BUILD_TEST)
     foreach(test_src ${Caffe2_XPU_TEST_SRCS})
       get_filename_component(test_name ${test_src} NAME_WE)
       add_executable(${test_name} "${test_src}")
-      torch_compile_options(${test_name})
       target_link_libraries(${test_name} torch_library gtest_main)
       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE})

docs/source/accelerator.md

@@ -40,7 +40,6 @@
     :nosignatures:
 
      empty_cache
-     get_memory_info
      max_memory_allocated
      max_memory_reserved
      memory_allocated

docs/source/conf.py

@@ -382,6 +382,20 @@ coverage_ignore_functions = [
     # torch.ao.quantization.backend_config.tensorrt
     "get_tensorrt_backend_config",
     "get_tensorrt_backend_config_dict",
+    # torch.ao.quantization.backend_config.utils
+    "entry_to_pretty_str",
+    "get_fused_module_classes",
+    "get_fuser_method_mapping",
+    "get_fusion_pattern_to_extra_inputs_getter",
+    "get_fusion_pattern_to_root_node_getter",
+    "get_module_to_qat_module",
+    "get_pattern_to_dtype_configs",
+    "get_pattern_to_input_type_to_index",
+    "get_qat_module_classes",
+    "get_root_module_to_quantized_reference_module",
+    "pattern_to_human_readable",
+    "remove_boolean_dispatch_from_name",
+    # torch.ao.quantization.backend_config.x86
     "get_x86_backend_config",
     # torch.ao.quantization.fuse_modules
     "fuse_known_modules",
@@ -412,6 +426,25 @@ coverage_ignore_functions = [
     "insert_observers_for_model",
     "prepare",
     "propagate_dtypes_for_known_nodes",
+    # torch.ao.quantization.fx.utils
+    "all_node_args_except_first",
+    "all_node_args_have_no_tensors",
+    "assert_and_get_unique_device",
+    "collect_producer_nodes",
+    "create_getattr_from_value",
+    "create_node_from_old_node_preserve_meta",
+    "get_custom_module_class_keys",
+    "get_linear_prepack_op_for_dtype",
+    "get_new_attr_name_with_prefix",
+    "get_non_observable_arg_indexes_and_types",
+    "get_qconv_prepack_op",
+    "get_skipped_module_name_and_classes",
+    "graph_module_from_producer_nodes",
+    "maybe_get_next_module",
+    "node_arg_is_bias",
+    "node_arg_is_weight",
+    "return_arg_list",
+    # torch.ao.quantization.pt2e.graph_utils
     "bfs_trace_with_node_process",
     "find_sequential_partitions",
     "get_equivalent_types",
@@ -827,10 +860,80 @@ coverage_ignore_functions = [
     "get_latency_of_one_partition",
     "get_latency_of_partitioned_graph",
     "get_partition_to_latency_mapping",
+    # torch.fx.experimental.proxy_tensor
+    "decompose",
+    "disable_autocast_cache",
+    "disable_proxy_modes_tracing",
+    "dispatch_trace",
+    "extract_val",
+    "fake_signature",
+    "fetch_sym_proxy",
+    "fetch_object_proxy",
+    "get_innermost_proxy_mode",
+    "get_isolated_graphmodule",
+    "get_proxy_slot",
+    "get_torch_dispatch_modes",
+    "has_proxy_slot",
+    "is_sym_node",
+    "maybe_handle_decomp",
+    "proxy_call",
+    "set_meta",
+    "set_original_aten_op",
+    "set_proxy_slot",
+    "snapshot_fake",
+    "thunkify",
+    "track_tensor",
+    "track_tensor_tree",
+    "wrap_key",
+    "wrapper_and_args_for_make_fx",
+    # torch.fx.experimental.recording
     "record_shapeenv_event",
     "replay_shape_env_events",
     "shape_env_check_state_equal",
+    # torch.fx.experimental.sym_node
+    "ceil_impl",
+    "floor_ceil_helper",
+    "floor_impl",
+    "method_to_operator",
+    "sympy_is_channels_last_contiguous_2d",
+    "sympy_is_channels_last_contiguous_3d",
+    "sympy_is_channels_last_strides_2d",
+    "sympy_is_channels_last_strides_3d",
+    "sympy_is_channels_last_strides_generic",
+    "sympy_is_contiguous",
+    "sympy_is_contiguous_generic",
+    "to_node",
+    "wrap_node",
     "sym_sqrt",
+    # torch.fx.experimental.symbolic_shapes
+    "bind_symbols",
+    "cast_symbool_to_symint_guardless",
+    "create_contiguous",
+    "error",
+    "eval_guards",
+    "eval_is_non_overlapping_and_dense",
+    "expect_true",
+    "find_symbol_binding_fx_nodes",
+    "free_symbols",
+    "free_unbacked_symbols",
+    "fx_placeholder_targets",
+    "fx_placeholder_vals",
+    "guard_bool",
+    "guard_float",
+    "guard_int",
+    "guard_scalar",
+    "has_hint",
+    "has_symbolic_sizes_strides",
+    "is_channels_last_contiguous_2d",
+    "is_channels_last_contiguous_3d",
+    "is_channels_last_strides_2d",
+    "is_channels_last_strides_3d",
+    "is_contiguous",
+    "is_non_overlapping_and_dense_indicator",
+    "is_nested_int",
+    "is_symbol_binding_fx_node",
+    "is_symbolic",
+    # torch.fx.experimental.unification.core
     "reify",
     # torch.fx.experimental.unification.match
     "edge",
@@ -868,6 +971,24 @@ coverage_ignore_functions = [
     "reverse_dict",
     # torch.fx.experimental.unification.multipledispatch.variadic
     "isvariadic",
+    # torch.fx.experimental.unification.unification_tools
+    "assoc",
+    "assoc_in",
+    "dissoc",
+    "first",
+    "get_in",
+    "getter",
+    "groupby",
+    "itemfilter",
+    "itemmap",
+    "keyfilter",
+    "keymap",
+    "merge",
+    "merge_with",
+    "update_in",
+    "valfilter",
+    "valmap",
+    # torch.fx.experimental.unification.utils
     "freeze",
     "hashable",
     "raises",
@@ -1308,8 +1429,319 @@ coverage_ignore_functions = [
     # torch.onnx.symbolic_opset7
     "max",
     "min",
+    # torch.onnx.symbolic_opset8
+    "addmm",
+    "bmm",
+    "empty",
+    "empty_like",
+    "flatten",
+    "full",
+    "full_like",
+    "gt",
+    "lt",
+    "matmul",
+    "mm",
+    "ones",
+    "ones_like",
+    "prelu",
+    "repeat",
+    "zeros",
+    "zeros_like",
+    # torch.onnx.symbolic_opset9
+    "abs",
+    "acos",
+    "adaptive_avg_pool1d",
+    "adaptive_avg_pool2d",
+    "adaptive_avg_pool3d",
+    "adaptive_max_pool1d",
+    "adaptive_max_pool2d",
+    "adaptive_max_pool3d",
+    "add",
+    "addcmul",
+    "addmm",
+    "alias",
+    "amax",
+    "amin",
+    "aminmax",
+    "arange",
+    "argmax",
+    "argmin",
+    "as_strided",
+    "as_tensor",
+    "asin",
+    "atan",
+    "atan2",
+    "avg_pool1d",
+    "avg_pool2d",
+    "avg_pool3d",
+    "baddbmm",
+    "batch_norm",
+    "bernoulli",
+    "bitwise_not",
+    "bitwise_or",
+    "bmm",
+    "broadcast_tensors",
+    "broadcast_to",
+    "bucketize",
+    "cat",
+    "cdist",
+    "ceil",
+    "clamp",
+    "clamp_max",
+    "clamp_min",
+    "clone",
+    "constant_pad_nd",
+    "contiguous",
+    "conv1d",
+    "conv2d",
+    "conv3d",
+    "conv_tbc",
+    "conv_transpose1d",
+    "conv_transpose2d",
+    "conv_transpose3d",
+    "convert_element_type",
+    "convolution",
+    "cos",
+    "cosine_similarity",
+    "cross",
+    "cumsum",
+    "detach",
     "dim",
+    "div",
+    "dot",
+    "dropout",
+    "elu",
+    "embedding",
+    "embedding_bag",
+    "empty",
+    "empty_like",
+    "eq",
+    "erf",
+    "exp",
+    "expand",
+    "expand_as",
+    "eye",
+    "fill",
+    "flatten",
+    "floor",
+    "floor_divide",
+    "floordiv",
+    "frobenius_norm",
+    "full",
+    "full_like",
+    "gather",
+    "ge",
+    "gelu",
+    "get_pool_ceil_padding",
+    "glu",
+    "group_norm",
+    "gru",
+    "gt",
+    "hann_window",
+    "hardshrink",
+    "hardsigmoid",
+    "hardswish",
+    "hardtanh",
+    "index",
+    "index_add",
+    "index_copy",
+    "index_fill",
+    "index_put",
+    "index_select",
+    "instance_norm",
+    "is_floating_point",
+    "is_pinned",
+    "isnan",
+    "item",
+    "kl_div",
+    "layer_norm",
+    "le",
+    "leaky_relu",
+    "lerp",
+    "lift",
+    "linalg_cross",
+    "linalg_matrix_norm",
+    "linalg_norm",
+    "linalg_vector_norm",
+    "linear",
+    "linspace",
+    "log",
+    "log10",
+    "log1p",
+    "log2",
+    "log_sigmoid",
+    "log_softmax",
+    "logical_and",
+    "logical_not",
+    "logical_or",
+    "logical_xor",
+    "logit",
+    "logsumexp",
+    "lstm",
+    "lstm_cell",
+    "lt",
+    "masked_fill",
+    "masked_fill_",
+    "matmul",
+    "max",
+    "max_pool1d",
+    "max_pool1d_with_indices",
+    "max_pool2d",
+    "max_pool2d_with_indices",
+    "max_pool3d",
+    "max_pool3d_with_indices",
+    "maximum",
+    "meshgrid",
+    "min",
+    "minimum",
+    "mish",
+    "mm",
+    "movedim",
+    "mse_loss",
+    "mul",
+    "multinomial",
+    "mv",
+    "narrow",
+    "native_layer_norm",
+    "ne",
+    "neg",
+    "new_empty",
+    "new_full",
+    "new_ones",
+    "new_zeros",
+    "nonzero",
+    "nonzero_numpy",
+    "noop_complex_operators",
+    "norm",
+    "numel",
+    "numpy_T",
+    "one_hot",
+    "ones",
+    "ones_like",
+    "onnx_placeholder",
+    "overload_by_arg_count",
+    "pad",
+    "pairwise_distance",
+    "permute",
+    "pixel_shuffle",
+    "pixel_unshuffle",
+    "pow",
+    "prelu",
+    "prim_constant",
+    "prim_constant_chunk",
+    "prim_constant_split",
+    "prim_data",
+    "prim_device",
+    "prim_dtype",
+    "prim_if",
+    "prim_layout",
+    "prim_list_construct",
+    "prim_list_unpack",
+    "prim_loop",
+    "prim_max",
+    "prim_min",
+    "prim_shape",
+    "prim_tolist",
+    "prim_tuple_construct",
+    "prim_type",
+    "prim_unchecked_cast",
+    "prim_uninitialized",
+    "rand",
+    "rand_like",
+    "randint",
+    "randint_like",
+    "randn",
+    "randn_like",
+    "reciprocal",
+    "reflection_pad",
+    "relu",
+    "relu6",
+    "remainder",
+    "repeat",
+    "repeat_interleave",
+    "replication_pad",
+    "reshape",
+    "reshape_as",
+    "rnn_relu",
+    "rnn_tanh",
+    "roll",
+    "rrelu",
+    "rsqrt",
+    "rsub",
+    "scalar_tensor",
+    "scatter",
+    "scatter_add",
+    "select",
+    "selu",
+    "sigmoid",
+    "sign",
+    "silu",
+    "sin",
+    "size",
+    "slice",
+    "softmax",
+    "softplus",
+    "softshrink",
+    "sort",
+    "split",
+    "split_with_sizes",
+    "sqrt",
+    "square",
+    "squeeze",
+    "stack",
+    "std",
+    "std_mean",
+    "sub",
+    "t",
+    "take",
+    "tan",
+    "tanh",
+    "tanhshrink",
+    "tensor",
+    "threshold",
+    "to",
+    "topk",
+    "transpose",
+    "true_divide",
+    "type_as",
+    "unbind",
+    "unfold",
+    "unsafe_chunk",
+    "unsafe_split",
+    "unsafe_split_with_sizes",
+    "unsqueeze",
+    "unsupported_complex_operators",
+    "unused",
+    "upsample_bilinear2d",
+    "upsample_linear1d",
+    "upsample_nearest1d",
+    "upsample_nearest2d",
+    "upsample_nearest3d",
+    "upsample_trilinear3d",
+    "var",
+    "var_mean",
+    "view",
+    "view_as",
+    "where",
+    "wrap_logical_op_with_cast_to",
+    "wrap_logical_op_with_negation",
+    "zero",
+    "zeros",
+    "zeros_like",
+    # torch.onnx.utils
+    "disable_apex_o2_state_dict_hook",
     "export",
+    "export_to_pretty_string",
+    "exporter_context",
+    "is_in_onnx_export",
+    "model_signature",
+    "register_custom_op_symbolic",
+    "select_model_mode_for_export",
+    "setup_onnx_logging",
+    "unconvertible_ops",
+    "unpack_quantized_tensor",
+    "warn_on_static_input_change",
+    # torch.onnx.verification
     "check_export_model_diff",
     "verify",
     "verify_aten_graph",
@@ -1400,6 +1832,32 @@ coverage_ignore_functions = [
     "noop_context_fn",
     "set_checkpoint_early_stop",
     "set_device_states",
+    # torch.utils.collect_env
+    "check_release_file",
+    "get_cachingallocator_config",
+    "get_clang_version",
+    "get_cmake_version",
+    "get_conda_packages",
+    "get_cpu_info",
+    "get_cuda_module_loading_config",
+    "get_cudnn_version",
+    "get_env_info",
+    "get_gcc_version",
+    "get_gpu_info",
+    "get_libc_version",
+    "get_lsb_version",
+    "get_mac_version",
+    "get_nvidia_driver_version",
+    "get_nvidia_smi",
+    "get_os",
+    "get_pip_packages",
+    "get_platform",
+    "get_pretty_env_info",
+    "get_python_platform",
+    "get_running_cuda_version",
+    "get_windows_version",
+    "is_xnnpack_available",
+    "pretty_str",
     # torch.utils.cpp_backtrace
     "get_cpp_backtrace",
     # torch.utils.cpp_extension
@@ -1463,6 +1921,52 @@ coverage_ignore_functions = [
     "apply_shuffle_seed",
     "apply_shuffle_settings",
     "get_all_graph_pipes",
+    # torch.utils.flop_counter
+    "addmm_flop",
+    "baddbmm_flop",
+    "bmm_flop",
+    "conv_backward_flop",
+    "conv_flop",
+    "conv_flop_count",
+    "convert_num_with_suffix",
+    "get_shape",
+    "get_suffix_str",
+    "mm_flop",
+    "normalize_tuple",
+    "register_flop_formula",
+    "sdpa_backward_flop",
+    "sdpa_backward_flop_count",
+    "sdpa_flop",
+    "sdpa_flop_count",
+    "shape_wrapper",
+    "transpose_shape",
+    # torch.utils.hipify.hipify_python
+    "add_dim3",
+    "compute_stats",
+    "extract_arguments",
+    "file_add_header",
+    "file_specific_replacement",
+    "find_bracket_group",
+    "find_closure_group",
+    "find_parentheses_group",
+    "fix_static_global_kernels",
+    "get_hip_file_path",
+    "hip_header_magic",
+    "hipify",
+    "is_caffe2_gpu_file",
+    "is_cusparse_file",
+    "is_out_of_place",
+    "is_pytorch_file",
+    "is_special_file",
+    "match_extensions",
+    "matched_files_iter",
+    "openf",
+    "preprocess_file_and_save_result",
+    "preprocessor",
+    "processKernelLaunches",
+    "replace_extern_shared",
+    "replace_math_functions",
+    "str2bool",
     # torch.utils.hooks
     "unserializable_hook",
     "warn_if_has_hooks",

docs/source/fx.experimental.md

@@ -12,37 +12,6 @@ These APIs are experimental and subject to change without notice.
 .. autoclass:: torch.fx.experimental.sym_node.DynamicInt
 ```
 
-## torch.fx.experimental.sym_node
-
-```{eval-rst}
-.. currentmodule:: torch.fx.experimental.sym_node
-```
-
-```{eval-rst}
-.. automodule:: torch.fx.experimental.sym_node
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    is_channels_last_contiguous_2d
-    is_channels_last_contiguous_3d
-    is_channels_last_strides_2d
-    is_channels_last_strides_3d
-    is_contiguous
-    is_non_overlapping_and_dense_indicator
-    method_to_operator
-    sympy_is_channels_last_contiguous_2d
-    sympy_is_channels_last_contiguous_3d
-    sympy_is_channels_last_strides_2d
-    sympy_is_channels_last_strides_3d
-    sympy_is_channels_last_strides_generic
-    sympy_is_contiguous
-    sympy_is_contiguous_generic
-```
-
 ## torch.fx.experimental.symbolic_shapes
 
 ```{eval-rst}
@@ -100,25 +69,6 @@ These APIs are experimental and subject to change without notice.
     rebind_unbacked
     resolve_unbacked_bindings
     is_accessor_node
-    cast_symbool_to_symint_guardless
-    create_contiguous
-    error
-    eval_guards
-    eval_is_non_overlapping_and_dense
-    find_symbol_binding_fx_nodes
-    free_symbols
-    free_unbacked_symbols
-    fx_placeholder_targets
-    fx_placeholder_vals
-    guard_bool
-    guard_float
-    guard_int
-    guard_scalar
-    has_hint
-    has_symbolic_sizes_strides
-    is_nested_int
-    is_symbol_binding_fx_node
-    is_symbolic
 ```
 
 ## torch.fx.experimental.proxy_tensor
@@ -141,46 +91,4 @@ These APIs are experimental and subject to change without notice.
     get_proxy_mode
     maybe_enable_thunkify
     maybe_disable_thunkify
-    decompose
-    disable_autocast_cache
-    disable_proxy_modes_tracing
-    extract_val
-    fake_signature
-    fetch_object_proxy
-    fetch_sym_proxy
-    has_proxy_slot
-    is_sym_node
-    maybe_handle_decomp
-    proxy_call
-    set_meta
-    set_original_aten_op
-    set_proxy_slot
-    snapshot_fake
 ```
-
-## torch.fx.experimental.unification.unification_tools
-
-```{eval-rst}
-.. currentmodule:: torch.fx.experimental.unification.unification_tools
-```
-
-```{eval-rst}
-.. automodule:: torch.fx.experimental.unification.unification_tools
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    assoc
-    assoc_in
-    dissoc
-    first
-    keyfilter
-    keymap
-    merge
-    merge_with
-    update_in
-    valfilter
-    valmap

docs/source/fx.md

@@ -1134,6 +1134,7 @@ The set of leaf modules can be customized by overriding
 .. py:module:: torch.fx.experimental.refinement_types
 .. py:module:: torch.fx.experimental.rewriter
 .. py:module:: torch.fx.experimental.schema_type_annotation
+.. py:module:: torch.fx.experimental.sym_node
 .. py:module:: torch.fx.experimental.unification.core
 .. py:module:: torch.fx.experimental.unification.dispatch
 .. py:module:: torch.fx.experimental.unification.match
@@ -1143,6 +1144,7 @@ The set of leaf modules can be customized by overriding
 .. py:module:: torch.fx.experimental.unification.multipledispatch.dispatcher
 .. py:module:: torch.fx.experimental.unification.multipledispatch.utils
 .. py:module:: torch.fx.experimental.unification.multipledispatch.variadic
+.. py:module:: torch.fx.experimental.unification.unification_tools
 .. py:module:: torch.fx.experimental.unification.utils
 .. py:module:: torch.fx.experimental.unification.variable
 .. py:module:: torch.fx.experimental.unify_refinements

docs/source/quantization-support.md

@@ -134,23 +134,6 @@ Quantization to work with this as well.
     ObservationType
 ```
 
-## torch.ao.quantization.backend_config.utils
-```{eval-rst}
-.. currentmodule:: torch.ao.quantization.backend_config.utils
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-    :template: classtemplate.rst
-
-    entry_to_pretty_str
-    pattern_to_human_readable
-    remove_boolean_dispatch_from_name
-
-```
-
 ## torch.ao.quantization.fx.custom_config
 
 This module contains a few CustomConfig classes that's used in both eager mode and FX graph mode quantization
@@ -171,30 +154,6 @@ This module contains a few CustomConfig classes that's used in both eager mode a
     StandaloneModuleConfigEntry
 ```
 
-## torch.ao.quantization.fx.utils
-
-```{eval-rst}
-.. currentmodule:: torch.ao.quantization.fx.utils
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-    :template: classtemplate.rst
-
-    all_node_args_except_first
-    all_node_args_have_no_tensors
-    collect_producer_nodes
-    create_getattr_from_value
-    create_node_from_old_node_preserve_meta
-    graph_module_from_producer_nodes
-    maybe_get_next_module
-    node_arg_is_bias
-    node_arg_is_weight
-    return_arg_list
-```
-
 ## torch.ao.quantization.quantizer
 
 ```{eval-rst}

docs/source/utils.md

@@ -19,91 +19,6 @@
     swap_tensors
 ```
 
-# torch.utils.collect_env
-```{eval-rst}
-.. automodule:: torch.utils.collect_env
-```
-
-```{eval-rst}
-.. currentmodule:: torch.utils.collect_env
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    check_release_file
-    is_xnnpack_available
-    pretty_str
-```
-
-# torch.utils.flop_counter
-```{eval-rst}
-.. automodule:: torch.utils.flop_counter
-```
-
-```{eval-rst}
-.. currentmodule:: torch.utils.flop_counter
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    baddbmm_flop
-    bmm_flop
-    conv_backward_flop
-    conv_flop
-    conv_flop_count
-    register_flop_formula
-    sdpa_backward_flop
-    sdpa_backward_flop_count
-    sdpa_flop
-    sdpa_flop_count
-    shape_wrapper
-```
-
-# torch.utils.hipify.hipify_python
-```{eval-rst}
-.. automodule:: torch.utils.hipify.hipify_python
-```
-
-```{eval-rst}
-.. currentmodule:: torch.utils.hipify.hipify_python
-```
-
-```{eval-rst}
-.. autosummary::
-    :toctree: generated
-    :nosignatures:
-
-    compute_stats
-    extract_arguments
-    file_add_header
-    file_specific_replacement
-    find_bracket_group
-    find_closure_group
-    find_parentheses_group
-    fix_static_global_kernels
-    hip_header_magic
-    hipify
-    is_caffe2_gpu_file
-    is_cusparse_file
-    is_out_of_place
-    is_pytorch_file
-    is_special_file
-    openf
-    preprocess_file_and_save_result
-    preprocessor
-    processKernelLaunches
-    replace_extern_shared
-    replace_math_functions
-    str2bool
-```
-
-
 <!-- This module needs to be documented. Adding here in the meantime
 for tracking purposes -->
 ```{eval-rst}
@@ -128,6 +43,7 @@ for tracking purposes -->
 .. py:module:: torch.utils.benchmark.utils.valgrind_wrapper.timer_interface
 .. py:module:: torch.utils.bundled_inputs
 .. py:module:: torch.utils.checkpoint
+.. py:module:: torch.utils.collect_env
 .. py:module:: torch.utils.cpp_backtrace
 .. py:module:: torch.utils.cpp_extension
 .. py:module:: torch.utils.data.backward_compatibility
@@ -164,8 +80,10 @@ for tracking purposes -->
 .. py:module:: torch.utils.data.sampler
 .. py:module:: torch.utils.dlpack
 .. py:module:: torch.utils.file_baton
+.. py:module:: torch.utils.flop_counter
 .. py:module:: torch.utils.hipify.constants
 .. py:module:: torch.utils.hipify.cuda_to_hip_mappings
+.. py:module:: torch.utils.hipify.hipify_python
 .. py:module:: torch.utils.hipify.version
 .. py:module:: torch.utils.hooks
 .. py:module:: torch.utils.jit.log_extract

pyproject.toml

@@ -172,9 +172,9 @@ ignore = [
     "SIM102", "SIM103", "SIM112", # flake8-simplify code styles
     "SIM105", # these ignores are from flake8-simplify. please fix or ignore with commented reason
     "SIM108", # SIM108 ignored because we prefer if-else-block instead of ternary expression
-    "SIM110", # Checks for for loops that can be replaced with a builtin function, like any or all.
+    "SIM110",
     "SIM114", # Combine `if` branches using logical `or` operator
-    "SIM115", # Checks for cases where files are opened without using a context manager.
+    "SIM115",
     "SIM116", # Disable Use a dictionary instead of consecutive `if` statements
     "SIM117",
     "SIM118",
@@ -184,6 +184,7 @@ ignore = [
     "TC006",
     # TODO: Remove Python-3.10 specific suppressions
     "B905",
+    "UP035",
 ]
 select = [
     "B",

setup.py

@@ -1646,7 +1646,8 @@ def main() -> None:
     mirror_files_into_torchgen()
     if RUN_BUILD_DEPS:
         build_deps()
-        mirror_inductor_external_kernels()
+
+    mirror_inductor_external_kernels()
 
     (
         ext_modules,

test/ao/sparsity/test_data_sparsifier.py

@@ -208,7 +208,7 @@ class _BaseDataSparsiferTestCase(TestCase):
         assert len(sparsifier1.data_groups) == len(sparsifier2.data_groups)
 
         state1 = state_dict1["state"]
-        for name in state1:
+        for name in state1.keys():
             # compare mask
             assert name in sparsifier2.state
             assert "mask" in sparsifier2.state[name]

test/ao/sparsity/test_sparsifier.py

@@ -119,7 +119,7 @@ class TestBaseSparsifier(TestCase):
         for idx in range(len(sparsifier0.groups)):
             mg0 = sparsifier0.groups[idx]
             mg1 = sparsifier1.groups[idx]
-            for key in mg0:
+            for key in mg0.keys():
                 assert key in mg1
                 if key == "module":
                     # We cannot compare modules as they are different

test/cpp/aoti_abi_check/CMakeLists.txt

@@ -10,8 +10,6 @@ set(AOTI_ABI_CHECK_TEST_SRCS
   ${AOTI_ABI_CHECK_TEST_ROOT}/main.cpp
   ${AOTI_ABI_CHECK_TEST_ROOT}/test_cast.cpp
   ${AOTI_ABI_CHECK_TEST_ROOT}/test_devicetype.cpp
-  ${AOTI_ABI_CHECK_TEST_ROOT}/test_dispatch.cpp
-  ${AOTI_ABI_CHECK_TEST_ROOT}/test_dispatch_v2.cpp
   ${AOTI_ABI_CHECK_TEST_ROOT}/test_dtype.cpp
   ${AOTI_ABI_CHECK_TEST_ROOT}/test_exception.cpp
   ${AOTI_ABI_CHECK_TEST_ROOT}/test_headeronlyarrayref.cpp

test/cpp/aoti_abi_check/test_dispatch.cpp

@@ -1,82 +0,0 @@
-#include <gtest/gtest.h>
-
-#include <torch/headeronly/core/Dispatch.h>
-#include <torch/headeronly/core/Dispatch_v2.h>
-
-// MY_PRIVATE_CHECK_SELECTIVE_BUILD is a prelude to case block. For
-// testing, we do nothing:
-#define MY_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type) /* empty */
-
-#define MY_PRIVATE_CASE_TYPE_USING_HINT(...) \
-  THO_PRIVATE_CASE_TYPE_USING_HINT_TMPL(     \
-      MY_PRIVATE_CHECK_SELECTIVE_BUILD, __VA_ARGS__)
-
-#define MY_DISPATCH_CASE(...) \
-  THO_DISPATCH_CASE_TMPL(MY_PRIVATE_CASE_TYPE_USING_HINT, __VA_ARGS__)
-
-// MY_RECORD_KERNEL_FUNCTION_DTYPE is a prelude to switch
-// statement. For testing, we just avoid unused variable warning:
-#define MY_RECORD_KERNEL_FUNCTION_DTYPE(DISPATCHNAME, ENUMTYPE) \
-  (void)DISPATCHNAME
-
-// MY_CHECK_NOT_IMPLEMENTED is called in switch default block. For
-// testing, we count case mismatches:
-#define MY_CHECK_NOT_IMPLEMENTED(...) default_count++
-
-#define MY_DISPATCH_SWITCH(...) \
-  THO_DISPATCH_SWITCH_TMPL(     \
-      MY_RECORD_KERNEL_FUNCTION_DTYPE, MY_CHECK_NOT_IMPLEMENTED, __VA_ARGS__)
-
-// MY_CASE_FUNCTION is called in a case block. For testing, we count
-// case matches and ensure that scalar_t/index_t type is defined:
-#define MY_CASE_FUNCTION \
-  [&] {                  \
-    count++;             \
-    scalar_t tmp;        \
-    (void)tmp;           \
-  }
-#define MY_INDEX_CASE_FUNCTION \
-  [&] {                        \
-    count++;                   \
-    index_t tmp;               \
-    (void)tmp;                 \
-  }
-
-#define DEFINE_ITEM(TYPE, SCALARTYPE) ScalarType::SCALARTYPE,
-
-#define MY_DISPATCH_V2(TYPE, NAME, BODY, ...) \
-  THO_DISPATCH_V2_TMPL(                       \
-      MY_DISPATCH_SWITCH,                     \
-      MY_DISPATCH_CASE,                       \
-      TYPE,                                   \
-      NAME,                                   \
-      AT_WRAP(BODY),                          \
-      __VA_ARGS__)
-
-#define TEST_DISPATCH_V2(NAME, EXPECTEDCOUNT, ...)                             \
-  TEST(TestDispatchV2, NAME) {                                                 \
-    using torch::headeronly::ScalarType;                                       \
-    using torch::headeronly::impl::ScalarTypeToCPPTypeT;                       \
-    int8_t total_count = 0;                                                    \
-    int8_t count = 0;                                                          \
-    int8_t default_count = 0;                                                  \
-    for (ScalarType t :                                                        \
-         {AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_ITEM)}) {       \
-      total_count++;                                                           \
-      MY_DISPATCH_V2(t, "test_my_dispatch_v2", MY_CASE_FUNCTION, __VA_ARGS__); \
-    }                                                                          \
-    EXPECT_EQ(count, EXPECTEDCOUNT);                                           \
-    EXPECT_EQ(default_count + count, total_count);                             \
-  }
-
-TEST_DISPATCH_V2(AT_FLOAT8_TYPES_, 5, AT_FLOAT8_TYPES);
-TEST_DISPATCH_V2(AT_INTEGRAL_TYPES_, 5, AT_INTEGRAL_TYPES);
-TEST_DISPATCH_V2(AT_FLOATING_TYPES_, 2, AT_FLOATING_TYPES);
-TEST_DISPATCH_V2(AT_BAREBONES_UNSIGNED_TYPES_, 3, AT_BAREBONES_UNSIGNED_TYPES);
-TEST_DISPATCH_V2(AT_INTEGRAL_TYPES_V2_, 8, AT_INTEGRAL_TYPES_V2);
-TEST_DISPATCH_V2(AT_COMPLEX_TYPES_, 2, AT_COMPLEX_TYPES);
-TEST_DISPATCH_V2(AT_QINT_TYPES_, 3, AT_QINT_TYPES);
-TEST_DISPATCH_V2(AT_ALL_TYPES_, 7, AT_ALL_TYPES);
-TEST_DISPATCH_V2(AT_ALL_TYPES_AND_COMPLEX_, 9, AT_ALL_TYPES_AND_COMPLEX);
-
-#undef DEFINE_ITEM

test/cpp/aoti_abi_check/test_dispatch_v2.cpp

@@ -1,45 +0,0 @@
-#include <gtest/gtest.h>
-#include <torch/headeronly/core/Dispatch_v2.h>
-#include <torch/headeronly/util/Exception.h>
-
-#define DEFINE_ITEM(TYPE, SCALARTYPE) ScalarType::SCALARTYPE,
-
-#define TEST_DISPATCH_V2(NAME, EXPECTEDCOUNT, ...)                       \
-  TEST(TestThoDispatchV2, NAME) {                                        \
-    using torch::headeronly::ScalarType;                                 \
-    using torch::headeronly::impl::ScalarTypeToCPPTypeT;                 \
-    int8_t total_count = 0;                                              \
-    int8_t count = 0;                                                    \
-    int8_t default_count = 0;                                            \
-    for (ScalarType t :                                                  \
-         {AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_ITEM)}) { \
-      total_count++;                                                     \
-      try {                                                              \
-        THO_DISPATCH_V2(                                                 \
-            t,                                                           \
-            "test_tho_dispatch_v2",                                      \
-            [&] {                                                        \
-              count++;                                                   \
-              scalar_t tmp;                                              \
-              (void)tmp;                                                 \
-            },                                                           \
-            __VA_ARGS__);                                                \
-      } catch (...) {                                                    \
-        default_count++; /* counts mismatches */                         \
-      }                                                                  \
-    }                                                                    \
-    EXPECT_EQ(count, EXPECTEDCOUNT);                                     \
-    EXPECT_EQ(default_count + count, total_count);                       \
-  }
-
-TEST_DISPATCH_V2(AT_FLOAT8_TYPES_, 5, AT_FLOAT8_TYPES);
-TEST_DISPATCH_V2(AT_INTEGRAL_TYPES_, 5, AT_INTEGRAL_TYPES);
-TEST_DISPATCH_V2(AT_FLOATING_TYPES_, 2, AT_FLOATING_TYPES);
-TEST_DISPATCH_V2(AT_BAREBONES_UNSIGNED_TYPES_, 3, AT_BAREBONES_UNSIGNED_TYPES);
-TEST_DISPATCH_V2(AT_INTEGRAL_TYPES_V2_, 8, AT_INTEGRAL_TYPES_V2);
-TEST_DISPATCH_V2(AT_COMPLEX_TYPES_, 2, AT_COMPLEX_TYPES);
-TEST_DISPATCH_V2(AT_QINT_TYPES_, 3, AT_QINT_TYPES);
-TEST_DISPATCH_V2(AT_ALL_TYPES_, 7, AT_ALL_TYPES);
-TEST_DISPATCH_V2(AT_ALL_TYPES_AND_COMPLEX_, 9, AT_ALL_TYPES_AND_COMPLEX);
-
-#undef DEFINE_ITEM

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/csrc/kernel.cpp

@@ -67,13 +67,13 @@ Tensor sgd_out_of_place(
 
 void boxed_sgd_out_of_place(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
   Tensor res = sgd_out_of_place(
-    torch::stable::detail::to<Tensor>(stack[0]),
-    torch::stable::detail::to<Tensor>(stack[1]),
-    float(torch::stable::detail::to<double>(stack[2])),
-    torch::stable::detail::to<double>(stack[3]),
-    torch::stable::detail::to<bool>(stack[4]));
+    to<Tensor>(stack[0]),
+    to<Tensor>(stack[1]),
+    float(to<double>(stack[2])),
+    to<double>(stack[3]),
+    to<bool>(stack[4]));
 
-  stack[0] = torch::stable::detail::from(res);
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY(libtorch_agnostic, m) {
@@ -89,8 +89,8 @@ Tensor identity(Tensor t) {
 }
 
 void boxed_identity(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = identity(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  Tensor res = identity(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -108,14 +108,14 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CPU, m) {
 Tensor my_abs(Tensor t) {
   const auto num_args = 1;
   StableIValue stack[num_args];
-  stack[0] = torch::stable::detail::from(t);
+  stack[0] = from(t);
   aoti_torch_call_dispatcher("aten::abs", "", stack);
-  return torch::stable::detail::to<Tensor>(stack[0]);
+  return to<Tensor>(stack[0]);
 }
 
 void boxed_my_abs(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_abs(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(tensor_res);
+  Tensor tensor_res = my_abs(to<Tensor>(stack[0]));
+  stack[0] = from(tensor_res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -132,21 +132,21 @@ Tensor my_ones_like(Tensor t, StableIValue device) {
 
   auto mf = aoti_torch_memory_format_contiguous_format();
 
-  stack[0] = torch::stable::detail::from(t);
-  stack[1] = torch::stable::detail::from(std::optional(t.scalar_type()));    // dtype
-  stack[2] = torch::stable::detail::from(std::nullopt);              // layout
-  stack[3] = torch::stable::detail::from(std::optional(device));     // device
-  stack[4] = torch::stable::detail::from(std::optional(false));      // pin_memory
-  stack[5] = torch::stable::detail::from(std::optional(mf));         // memory_format
+  stack[0] = from(t);
+  stack[1] = from(std::optional(t.scalar_type()));    // dtype
+  stack[2] = from(std::nullopt);              // layout
+  stack[3] = from(std::optional(device));     // device
+  stack[4] = from(std::optional(false));      // pin_memory
+  stack[5] = from(std::optional(mf));         // memory_format
 
   aoti_torch_call_dispatcher("aten::ones_like", "", stack);
 
-  return torch::stable::detail::to<Tensor>(stack[0]);
+  return to<Tensor>(stack[0]);
 }
 
 void boxed_my_ones_like(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = my_ones_like(torch::stable::detail::to<Tensor>(stack[0]), stack[1]);
-  stack[0] = torch::stable::detail::from(res);
+  Tensor res = my_ones_like(to<Tensor>(stack[0]), stack[1]);
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -159,28 +159,28 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
 
 std::tuple<Tensor, Tensor, bool> exp_neg_is_leaf(Tensor t1, Tensor t2, Tensor t3) {
   StableIValue stack_exp[1];
-  stack_exp[0] = torch::stable::detail::from(t1);
+  stack_exp[0] = from(t1);
   aoti_torch_call_dispatcher("aten::exp", "", stack_exp);
 
   StableIValue stack_neg[1];
-  stack_neg[0] = torch::stable::detail::from(t2);
+  stack_neg[0] = from(t2);
   aoti_torch_call_dispatcher("aten::neg", "", stack_neg);
 
   StableIValue stack_is_leaf[1];
-  stack_is_leaf[0] = torch::stable::detail::from(t3);
+  stack_is_leaf[0] = from(t3);
   aoti_torch_call_dispatcher("aten::is_leaf", "", stack_is_leaf);
 
   return std::make_tuple(
-    torch::stable::detail::to<Tensor>(stack_exp[0]),
-    torch::stable::detail::to<Tensor>(stack_neg[0]),
-    torch::stable::detail::to<bool>(stack_is_leaf[0]));
+    to<Tensor>(stack_exp[0]),
+    to<Tensor>(stack_neg[0]),
+    to<bool>(stack_is_leaf[0]));
 }
 
 void boxed_exp_neg_is_leaf(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto tuple = exp_neg_is_leaf(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]), torch::stable::detail::to<Tensor>(stack[2]));
-  stack[0] = torch::stable::detail::from(std::get<0>(tuple));
-  stack[1] = torch::stable::detail::from(std::get<1>(tuple));
-  stack[2] = torch::stable::detail::from(std::get<2>(tuple));
+  auto tuple = exp_neg_is_leaf(to<Tensor>(stack[0]), to<Tensor>(stack[1]), to<Tensor>(stack[2]));
+  stack[0] = from(std::get<0>(tuple));
+  stack[1] = from(std::get<1>(tuple));
+  stack[2] = from(std::get<2>(tuple));
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -193,15 +193,15 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
 
 Tensor neg_exp(Tensor t) {
   StableIValue stack[1];
-  stack[0] = torch::stable::detail::from(t);
+  stack[0] = from(t);
   aoti_torch_call_dispatcher("aten::exp", "", stack);
   aoti_torch_call_dispatcher("aten::neg", "", stack);
-  return torch::stable::detail::to<Tensor>(stack[0]);
+  return to<Tensor>(stack[0]);
 }
 
 void boxed_neg_exp(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = neg_exp(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  Tensor res = neg_exp(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -214,10 +214,10 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
 
 Tensor divide_neg_exp(Tensor t) {
   StableIValue stack_neg[1];
-  stack_neg[0] = torch::stable::detail::from(t);
+  stack_neg[0] = from(t);
 
   StableIValue stack_exp[1];
-  stack_exp[0] = torch::stable::detail::from(t);
+  stack_exp[0] = from(t);
   aoti_torch_call_dispatcher("aten::exp", "", stack_exp);
   aoti_torch_call_dispatcher("aten::neg", "", stack_neg);
 
@@ -225,12 +225,12 @@ Tensor divide_neg_exp(Tensor t) {
   stack_div[0] = stack_neg[0];
   stack_div[1] = stack_exp[0];
   aoti_torch_call_dispatcher("aten::divide", "Tensor", stack_div);
-  return torch::stable::detail::to<Tensor>(stack_div[0]);
+  return to<Tensor>(stack_div[0]);
 }
 
 void boxed_divide_neg_exp(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor res = divide_neg_exp(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  Tensor res = divide_neg_exp(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -246,8 +246,8 @@ bool is_contiguous(Tensor t) {
 }
 
 void boxed_is_contiguous(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  bool res = is_contiguous(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  bool res = is_contiguous(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -263,9 +263,9 @@ Tensor my_transpose(Tensor t, int64_t dim0, int64_t dim1) {
 }
 
 void boxed_my_transpose(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_transpose(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<int64_t>(stack[1]), torch::stable::detail::to<int64_t>(stack[2]));
+  auto res = my_transpose(to<Tensor>(stack[0]), to<int64_t>(stack[1]), to<int64_t>(stack[2]));
 
-  stack[0] = torch::stable::detail::from(res);
+  stack[0] = from(res);
 }
 
 Tensor my_empty_like(Tensor t) {
@@ -273,8 +273,8 @@ Tensor my_empty_like(Tensor t) {
 }
 
 void boxed_empty_like(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_empty_like(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_empty_like(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 bool my_is_cpu(Tensor t) {
@@ -283,8 +283,8 @@ bool my_is_cpu(Tensor t) {
 
 
 void boxed_my_is_cpu(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_is_cpu(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_is_cpu(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor fill_infinity(Tensor t) {
@@ -296,8 +296,8 @@ void boxed_fill_infinity(
     StableIValue* stack,
     uint64_t num_args,
     uint64_t num_outputs) {
-  auto res = fill_infinity(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = fill_infinity(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_pad(Tensor t) {
@@ -310,8 +310,8 @@ void boxed_my_pad(
     StableIValue* stack,
     uint64_t num_args,
     uint64_t num_outputs) {
-  auto res = my_pad(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_pad(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_narrow(Tensor t, int64_t dim, int64_t start, int64_t length) {
@@ -323,11 +323,11 @@ void boxed_my_narrow(
     uint64_t num_args,
     uint64_t num_outputs) {
   auto res = my_narrow(
-      torch::stable::detail::to<Tensor>(stack[0]),
-      torch::stable::detail::to<int64_t>(stack[1]),
-      torch::stable::detail::to<int64_t>(stack[2]),
-      torch::stable::detail::to<int64_t>(stack[3]));
-  stack[0] = torch::stable::detail::from(res);
+      to<Tensor>(stack[0]),
+      to<int64_t>(stack[1]),
+      to<int64_t>(stack[2]),
+      to<int64_t>(stack[3]));
+  stack[0] = from(res);
 }
 
 Tensor my_new_empty_dtype_variant(Tensor t) {
@@ -342,8 +342,8 @@ Tensor my_new_empty_dtype_variant(Tensor t) {
 }
 
 void boxed_my_new_empty_dtype_variant(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_new_empty_dtype_variant(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_new_empty_dtype_variant(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_new_zeros_dtype_variant(Tensor t) {
@@ -352,8 +352,8 @@ Tensor my_new_zeros_dtype_variant(Tensor t) {
 }
 
 void boxed_my_new_zeros_dtype_variant(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_new_zeros_dtype_variant(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_new_zeros_dtype_variant(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_copy_(Tensor dst, Tensor src, bool non_blocking) {
@@ -361,8 +361,8 @@ Tensor my_copy_(Tensor dst, Tensor src, bool non_blocking) {
 }
 
 void boxed_my_copy_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_copy_(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]), torch::stable::detail::to<bool>(stack[2]));
-  stack[0] = torch::stable::detail::from(tensor_res);
+  Tensor tensor_res = my_copy_(to<Tensor>(stack[0]), to<Tensor>(stack[1]), to<bool>(stack[2]));
+  stack[0] = from(tensor_res);
 }
 
 Tensor my_clone(Tensor t) {
@@ -370,8 +370,8 @@ Tensor my_clone(Tensor t) {
 }
 
 void boxed_my_clone(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  Tensor tensor_res = my_clone(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(tensor_res);
+  Tensor tensor_res = my_clone(to<Tensor>(stack[0]));
+  stack[0] = from(tensor_res);
 }
 
 
@@ -408,8 +408,8 @@ Tensor my_zero_(Tensor t) {
 }
 
 void boxed_my_zero_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_zero_(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_zero_(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_amax(Tensor t) {
@@ -417,8 +417,8 @@ Tensor my_amax(Tensor t) {
 }
 
 void boxed_my_amax(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_amax(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_amax(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 Tensor my_amax_vec(Tensor t) {
@@ -426,8 +426,8 @@ Tensor my_amax_vec(Tensor t) {
 }
 
 void boxed_my_amax_vec(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = my_amax_vec(torch::stable::detail::to<Tensor>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  auto res = my_amax_vec(to<Tensor>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -464,8 +464,8 @@ void boxed_test_default_constructor(
     StableIValue* stack,
     uint64_t num_args,
     uint64_t num_outputs) {
-  bool res = test_default_constructor(torch::stable::detail::to<bool>(stack[0]));
-  stack[0] = torch::stable::detail::from(res);
+  bool res = test_default_constructor(to<bool>(stack[0]));
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -478,56 +478,6 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
   m.impl("my_amax_vec", &boxed_my_amax_vec);
 }
 
-std::vector<Tensor> my__foreach_mul(torch::headeronly::HeaderOnlyArrayRef<Tensor> self, torch::headeronly::HeaderOnlyArrayRef<Tensor> other) {
-  std::array<StableIValue, 2> stack = {torch::stable::detail::from(self), torch::stable::detail::from(other)};
-  aoti_torch_call_dispatcher("aten::_foreach_mul", "List", stack.data());
-  return torch::stable::detail::to<std::vector<Tensor>>(stack[0]);
-}
-
-void boxed_my__foreach_mul(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  // Why is the following NOT torch::stable::detail::to<HeaderOnlyArrayRef<Tensor>>(stack[0])? Because calling `to`
-  // on a StableIValue means that the result is owning its underlying data now! HeaderOnlyArrayRef
-  // is not owning, so it cannot safely steward the result of the torch::stable::detail::to<>.
-  auto res = my__foreach_mul(torch::stable::detail::to<std::vector<Tensor>>(stack[0]), torch::stable::detail::to<std::vector<Tensor>>(stack[1]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
-void my__foreach_mul_(torch::headeronly::HeaderOnlyArrayRef<Tensor> self, torch::headeronly::HeaderOnlyArrayRef<Tensor> other) {
-  std::array<StableIValue, 2> stack = {torch::stable::detail::from(self), torch::stable::detail::from(other)};
-  aoti_torch_call_dispatcher("aten::_foreach_mul_", "List", stack.data());
-}
-
-void boxed_my__foreach_mul_(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  my__foreach_mul_(torch::stable::detail::to<std::vector<Tensor>>(stack[0]), torch::stable::detail::to<std::vector<Tensor>>(stack[1]));
-}
-
-std::vector<Tensor> make_tensor_clones_and_call_foreach(Tensor t1, Tensor t2) {
-  // This function tests that my__foreach_mul can take in std::initializer_lists
-  // in addition to std::vectors.
-  Tensor t1_1 = my_clone(t1);
-  Tensor t1_2 = my_clone(t1);
-  Tensor t2_1 = my_clone(t2);
-  Tensor t2_2 = my_clone(t2);
-  return my__foreach_mul({t1_1, t2_1}, {t1_2, t2_2});
-}
-
-void boxed_make_tensor_clones_and_call_foreach(StableIValue* stack, uint64_t num_args, uint64_t num_outputs) {
-  auto res = make_tensor_clones_and_call_foreach(torch::stable::detail::to<Tensor>(stack[0]), torch::stable::detail::to<Tensor>(stack[1]));
-  stack[0] = torch::stable::detail::from(res);
-}
-
-STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
-  m.def("my__foreach_mul(Tensor[] self, Tensor[] other) -> Tensor[]");
-  m.def("my__foreach_mul_(Tensor(a!)[] self, Tensor[] other) -> ()");
-  m.def("make_tensor_clones_and_call_foreach(Tensor t1, Tensor t2) -> Tensor[]");
-}
-
-STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
-  m.impl("my__foreach_mul", &boxed_my__foreach_mul);
-  m.impl("my__foreach_mul_", &boxed_my__foreach_mul_);
-  m.impl("make_tensor_clones_and_call_foreach", &boxed_make_tensor_clones_and_call_foreach);
-}
-
 // Test functions for torch::stable::accelerator APIs
 
 #ifdef LAE_USE_CUDA
@@ -550,8 +500,8 @@ void boxed_test_device_guard(
     StableIValue* stack,
     uint64_t num_args,
     uint64_t num_outputs) {
-  int res = test_device_guard(static_cast<int64_t>(torch::stable::detail::to<int64_t>(stack[0])));
-  stack[0] = torch::stable::detail::from(res);
+  int res = test_device_guard(static_cast<int64_t>(to<int64_t>(stack[0])));
+  stack[0] = from(res);
 }
 
 int64_t test_device_guard_set_index() {
@@ -570,7 +520,7 @@ void boxed_test_device_guard_set_index(
     uint64_t num_args,
     uint64_t num_outputs) {
   int64_t res = test_device_guard_set_index();
-  stack[0] = torch::stable::detail::from(res);
+  stack[0] = from(res);
 }
 
 int64_t test_stream(int32_t device_index) {
@@ -586,8 +536,8 @@ void boxed_test_stream(
     StableIValue* stack,
     uint64_t num_args,
     uint64_t num_outputs) {
-  int64_t res = test_stream(static_cast<int64_t>(torch::stable::detail::to<int64_t>(stack[0])));
-  stack[0] = torch::stable::detail::from(res);
+  int64_t res = test_stream(static_cast<int64_t>(to<int64_t>(stack[0])));
+  stack[0] = from(res);
 }
 
 int64_t test_get_current_device_index() {
@@ -599,7 +549,7 @@ void boxed_test_get_current_device_index(
     uint64_t num_args,
     uint64_t num_outputs) {
   int64_t res = test_get_current_device_index();
-  stack[0] = torch::stable::detail::from(res);
+  stack[0] = from(res);
 }
 
 STABLE_TORCH_LIBRARY_FRAGMENT(libtorch_agnostic, m) {
@@ -615,5 +565,4 @@ STABLE_TORCH_LIBRARY_IMPL(libtorch_agnostic, CompositeExplicitAutograd, m) {
   m.impl("test_stream", &boxed_test_stream);
   m.impl("test_get_current_device_index", &boxed_test_get_current_device_index);
 }
-
 #endif // LAE_USE_CUDA

test/cpp_extensions/libtorch_agnostic_extension/libtorch_agnostic/ops.py

@@ -333,45 +333,3 @@ def my_new_zeros_dtype_variant(t) -> Tensor:
     Returns: New zeros tensor
     """
     return torch.ops.libtorch_agnostic.my_new_zeros_dtype_variant.default(t)
-
-
-def my__foreach_mul_(tensors, others) -> ():
-    """
-    Updates tensors to be the result of pointwise multiplying with others.
-
-    Args:
-        tensors: list of tensors
-        others: list of tensors (with the same corresponding shapes as tensors)
-
-    Returns: nothing, tensors is updated in place.
-    """
-    torch.ops.libtorch_agnostic.my__foreach_mul_.default(tensors, others)
-
-
-def my__foreach_mul(tensors, others) -> list[Tensor]:
-    """
-    Returns a list of tensors that are the results of pointwise multiplying
-    tensors and others.
-
-    Args:
-        tensors: list of tensors
-        others: list of tensors (with the same corresponding shapes as tensors)
-
-    Returns: list of multiplied tensors
-    """
-    return torch.ops.libtorch_agnostic.my__foreach_mul.default(tensors, others)
-
-
-def make_tensor_clones_and_call_foreach(t1, t2) -> list[Tensor]:
-    """
-    Returns a list of 2 tensors corresponding to the square of the inputs.
-
-    Args:
-        t1: Tensor
-        t2: Tensor
-
-    Returns: list of [t1^2, t2^2]
-    """
-    return torch.ops.libtorch_agnostic.make_tensor_clones_and_call_foreach.default(
-        t1, t2
-    )

test/cpp_extensions/libtorch_agnostic_extension/test/test_libtorch_agnostic.py

@@ -367,57 +367,6 @@ if not IS_WINDOWS:
             self.assertNotEqual(result.data_ptr(), expected.data_ptr())
             self.assertEqual(result.stride(), expected.stride())
 
-        def test_my__foreach_mul_(self, device):
-            import libtorch_agnostic
-
-            N = 5
-            tensors = [torch.rand(32, 16, device=device) for _ in range(N)]
-            tensors_c = [t.clone() for t in tensors]
-            others = [torch.rand(32, 16, device=device) for _ in range(N)]
-
-            libtorch_agnostic.ops.my__foreach_mul_(tensors, others)
-            expected_values = torch._foreach_mul(tensors_c, others)
-
-            for tensor_t, expected_t in zip(tensors, expected_values):
-                self.assertEqual(tensor_t, expected_t)
-
-        def test_my__foreach_mul(self, device):
-            import libtorch_agnostic
-
-            N = 5
-            tensors = [torch.rand(32, 16, device=device) for _ in range(N)]
-            others = [torch.rand(32, 16, device=device) for _ in range(N)]
-
-            result = libtorch_agnostic.ops.my__foreach_mul(tensors, others)
-            expected = torch._foreach_mul(tensors, others)
-
-            for result_t, expected_t in zip(result, expected):
-                self.assertEqual(result_t, expected_t)
-
-            def _make_cuda_tensors(prior_mem):
-                cuda_res = libtorch_agnostic.ops.my__foreach_mul(tensors, others)
-                self.assertGreater(torch.cuda.memory_allocated(device), prior_mem)
-
-                expected = torch._foreach_mul(tensors, others)
-                for result_t, expected_t in zip(cuda_res, expected):
-                    self.assertEqual(result_t, expected_t)
-
-            if tensors[0].is_cuda:
-                init_mem = torch.cuda.memory_allocated(device)
-                for _ in range(3):
-                    _make_cuda_tensors(init_mem)
-                    curr_mem = torch.cuda.memory_allocated(device)
-                    self.assertEqual(curr_mem, init_mem)
-
-        def test_make_tensor_clones_and_call_foreach(self, device):
-            import libtorch_agnostic
-
-            t1 = torch.rand(2, 5, device=device)
-            t2 = torch.rand(3, 4, device=device)
-            result = libtorch_agnostic.ops.make_tensor_clones_and_call_foreach(t1, t2)
-            self.assertEqual(result[0], t1 * t1)
-            self.assertEqual(result[1], t2 * t2)
-
     instantiate_device_type_tests(TestLibtorchAgnostic, globals(), except_for=None)
 
 if __name__ == "__main__":

test/custom_backend/test_custom_backend.py

@@ -1,5 +1,6 @@
 # Owner(s): ["module: unknown"]
 
+import os
 import tempfile
 
 from backend import get_custom_backend_library_path, Model, to_custom_backend
@@ -40,11 +41,14 @@ class TestCustomBackend(TestCase):
         self.test_execute()
 
         # Save and load.
-        with tempfile.NamedTemporaryFile() as f:
+        f = tempfile.NamedTemporaryFile(delete=False)
+        try:
             f.close()
             torch.jit.save(self.model, f.name)
             loaded = torch.jit.load(f.name)
-            self.model = loaded
+        finally:
+            os.unlink(f.name)
+        self.model = loaded
 
         # Test execution again.
         self.test_execute()

test/custom_operator/test_custom_ops.py

@@ -1,5 +1,6 @@
 # Owner(s): ["module: unknown"]
 
+import os.path
 import sys
 import tempfile
 import unittest
@@ -143,13 +144,16 @@ def forward(self, arg0_1):
         # Ideally we would like to not have to manually delete the file, but NamedTemporaryFile
         # opens the file, and it cannot be opened multiple times in Windows. To support Windows,
         # close the file after creation and try to remove it manually.
-        with tempfile.NamedTemporaryFile() as file:
+        file = tempfile.NamedTemporaryFile(delete=False)
+        try:
             file.close()
             model.save(file.name)
             loaded = torch.jit.load(file.name)
+        finally:
+            os.unlink(file.name)
 
-            output = loaded.forward(torch.ones(5))
-            self.assertTrue(output.allclose(torch.ones(5) + 1))
+        output = loaded.forward(torch.ones(5))
+        self.assertTrue(output.allclose(torch.ones(5) + 1))
 
 
 if __name__ == "__main__":

test/distributed/_composable/fsdp/test_fully_shard_logging.py

@@ -1,7 +1,7 @@
 # Owner(s): ["module: fsdp"]
 import functools
 import os
-import unittest
+import unittest.mock
 
 import torch.distributed as dist
 from torch._dynamo.test_case import run_tests
@@ -37,9 +37,9 @@ import torch
 import torch.distributed as dist
 import torch.nn as nn
 from torch.distributed.fsdp import fully_shard
-logger = logging.getLogger("torch.distributed.fsdp.fully_shard")
+logger = logging.getLogger("torch.distributed._composable.fsdp")
 logger.setLevel(logging.DEBUG)
-device = '{device_type.type}'
+device = {device_type.type}
 torch.manual_seed(0)
 model = nn.Sequential(*[nn.Linear(4, 4, device=device, bias=False) for _ in range(2)])
 for layer in model:

test/distributed/_composable/test_replicate_with_fsdp.py

@@ -76,7 +76,7 @@ class ReplicateTest(MultiProcessTestCase):
             store=dist.FileStore(self.file_name, self.world_size),
         )
 
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_replicate_transformer(self):
         """
         This tests that replicate works on a transformer model with fully_shard and replicate layers
@@ -126,7 +126,7 @@ class ReplicateTest(MultiProcessTestCase):
                 for parameter in layer.parameters():
                     self.assertEqual(parameter.placements, (Shard(dim=0),))
 
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_replicate_transformer_managed_modules(self):
         """
         This tests that replicate managed modules works properly. In this test we use a Transformer Module with 3 layers,
@@ -178,7 +178,7 @@ class ReplicateTest(MultiProcessTestCase):
         replicate_model = replicate(replicate_model)
         self.assertEqual(len(_get_managed_modules((replicate_model,))), 21)
 
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_replicate_tp_device_mesh(self):
         """
         This tests that a user can pass in a device mesh to replicate a module
@@ -206,7 +206,7 @@ class ReplicateTest(MultiProcessTestCase):
                 self.assertEqual(parameter.device_mesh.shape, (2,))
                 self.assertEqual(parameter.placements, (Replicate(),))
 
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_train_replicate_fsdp(self):
         """
         Tests that replicate_model has the same behavior as original model when training
@@ -253,7 +253,7 @@ class ReplicateTest(MultiProcessTestCase):
             self.assertEqual(replicate_loss, loss)
             check_sharded_parity(self, model, replicate_model)
 
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_train_parity_2d_mlp(self):
         """
         Verifies when a device mesh is passed in, the model has the same behavior as the original model when training

test/distributed/_tools/test_sac_ilp.py

@@ -80,7 +80,7 @@ class TestSACILP(TestCase):
             # postprocessing due to the fact that for ModTracker, the post backward hook
             # is not being called for modules whose inputs don't require gradients
             # TODO: fix this in ModTracker and ensure it does not lead to any perf regression
-            if _ModState.POST_BW not in mod_stats.snapshots:
+            if _ModState.POST_BW not in mod_stats.snapshots.keys():
                 mod_stats.snapshots.setdefault(_ModState.POST_BW, []).append(
                     copy.deepcopy(last_snapshot)
                 )

test/distributed/argparse_util_test.py

@@ -16,7 +16,7 @@ from torch.distributed.argparse_util import check_env, env
 class ArgParseUtilTest(unittest.TestCase):
     def setUp(self):
         # remove any lingering environment variables
-        for e in os.environ.keys():  # noqa: SIM118
+        for e in os.environ.keys():
             if e.startswith("PET_"):
                 del os.environ[e]
 

test/distributed/checkpoint/_experimental/test_staging.py

@@ -207,7 +207,7 @@ class TestDefaultStager(TestCase):
         for i, result in enumerate(staged_results):
             self.assertIsInstance(result, dict)
             # Verify the result contains the expected keys
-            for key in state_dicts[i]:
+            for key in state_dicts[i].keys():
                 self.assertIn(key, result)
 
         stager.close()

test/distributed/checkpoint/test_dtensor_resharding.py

@@ -299,7 +299,7 @@ class TestDTensorReshardMeshChange(DTensorTestBase):
 
     @with_comms
     @with_temp_dir
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_dtensor_checkpoint_with_uneven_shards(self) -> None:
         """
         Saving a dtensor with uneven shards.
@@ -436,7 +436,6 @@ class TestCheckpointableReshard(DTensorTestBase):
 
     @with_comms
     @with_temp_dir
-    @skip_if_lt_x_gpu(4)
     def test_uneven_reshard_with_checkpointable_api(self) -> None:
         """
         Saves a 1d distributed tensor that has shards with uneven sizes using Checkpointable API.
@@ -499,7 +498,6 @@ class TestCheckpointableReshard(DTensorTestBase):
 
     @with_comms
     @with_temp_dir
-    @skip_if_lt_x_gpu(4)
     def test_uneven_reshard_with_dtensor_shards_wrapper_api(self) -> None:
         """
         Saves a 1d distributed tensor that has shards with uneven sizes using Checkpointable API.

test/distributed/checkpoint/test_hf_safetensor_e2e.py

@@ -60,7 +60,7 @@ class TestSingleRankSaveLoad(TestCase):
         self.assertEqual(
             sorted(state_dict_to_save.keys()), sorted(state_dict_loaded.keys())
         )
-        for key in state_dict_to_save:
+        for key in state_dict_to_save.keys():
             self.assertTrue(
                 torch.equal(state_dict_to_save[key], state_dict_loaded[key])
             )
@@ -89,7 +89,7 @@ class TestSingleRankSaveLoad(TestCase):
         self.assertEqual(
             sorted(state_dict_to_save.keys()), sorted(state_dict_to_load.keys())
         )
-        for key in state_dict_to_save:
+        for key in state_dict_to_save.keys():
             self.assertTrue(
                 torch.equal(state_dict_to_save[key], state_dict_to_load[key])
             )
@@ -116,7 +116,7 @@ class TestSingleRankSaveLoad(TestCase):
         self.assertEqual(
             sorted(state_dict_to_save.keys()), sorted(state_dict_loaded.keys())
         )
-        for key in state_dict_to_save:
+        for key in state_dict_to_save.keys():
             self.assertTrue(
                 torch.equal(state_dict_to_save[key], state_dict_loaded[key])
             )
@@ -156,7 +156,7 @@ class TestSingleRankSaveLoad(TestCase):
         self.assertEqual(
             sorted(state_dict_to_save.keys()), sorted(state_dict_to_load.keys())
         )
-        for key in state_dict_to_save:
+        for key in state_dict_to_save.keys():
             self.assertTrue(
                 torch.equal(state_dict_to_save[key], state_dict_to_load[key])
             )

test/distributed/checkpoint/test_planner.py

@@ -18,7 +18,6 @@ from torch.distributed.checkpoint._dedup_save_plans import dedup_save_plans
 from torch.distributed.checkpoint.api import CheckpointException
 from torch.distributed.checkpoint.default_planner import (
     _create_default_local_metadata,
-    _validate_global_plan,
     create_default_global_save_plan,
     create_default_local_load_plan,
     create_default_local_save_plan,
@@ -29,7 +28,6 @@ from torch.distributed.checkpoint.filesystem import CURRENT_DCP_VERSION
 from torch.distributed.checkpoint.metadata import (
     BytesStorageMetadata,
     ChunkStorageMetadata,
-    Metadata,
     MetadataIndex,
     TensorProperties,
     TensorStorageMetadata,
@@ -562,32 +560,6 @@ class TestPlannerHelpers(TestCase):
         self.assertTrue(_compare_save_plans(plan2, plan2))
 
 
-class TestValidateGlobalPlan(TestCase):
-    def _make_metadata(self, chunks, size):
-        storage = TensorStorageMetadata(
-            properties=TensorProperties(dtype=torch.float32),
-            size=torch.Size(size),
-            chunks=chunks,
-        )
-        return Metadata(state_dict_metadata={"param": storage})
-
-    def test_non_overlapping_chunks(self):
-        chunks = [
-            ChunkStorageMetadata(offsets=torch.Size([i]), sizes=torch.Size([1]))
-            for i in range(4)
-        ]
-        metadata = self._make_metadata(chunks, [4])
-        self.assertTrue(_validate_global_plan([SavePlan([])], metadata))
-
-    def test_detect_overlapping_chunks(self):
-        chunks = [
-            ChunkStorageMetadata(offsets=torch.Size([0]), sizes=torch.Size([2])),
-            ChunkStorageMetadata(offsets=torch.Size([1]), sizes=torch.Size([2])),
-        ]
-        metadata = self._make_metadata(chunks, [4])
-        self.assertFalse(_validate_global_plan([SavePlan([])], metadata))
-
-
 class TestLoadPlanner(TestCase):
     @with_temp_dir
     def test_strict(self):

test/distributed/checkpoint/test_state_dict.py

@@ -769,7 +769,7 @@ class TestStateDict(DTensorTestBase, VerifyStateDictMixin):
         model_state_dict3 = copy.deepcopy(model_state_dict3)
         self.assertEqual(len(model_state_dict2), 2)
         self.assertEqual(len(model_state_dict3), 2)
-        for key in model_state_dict3:
+        for key in model_state_dict3.keys():
             full_fqn = f"l.{key}"
             value1 = model_state_dict1[full_fqn]
             value2 = model_state_dict2[full_fqn]
@@ -886,7 +886,7 @@ class TestStateDict(DTensorTestBase, VerifyStateDictMixin):
             self.assertEqual(cpu_model_value, meta_model_value)
 
     @with_comms
-    @skip_if_lt_x_gpu(4)
+    @skip_if_lt_x_gpu(2)
     def test_setting_meta_device_model_broadcasting_and_memory(self) -> None:
         # This test verifies that we can set model state dict by a meta device model
         # With the correlated changes in state_dict, meta device model should be accepted

test/distributed/fsdp/test_fsdp_misc.py

@@ -479,7 +479,6 @@ class TestFSDPMiscMultiProcess(FSDPTest):
                     for (n, p), (n_prev, p_prev) in zip(
                         fsdp_overlap.named_parameters(), fsdp_overlap_prev_params
                     ):
-                        self.assertEqual(n, n_prev)
                         self.assertNotEqual(
                             p,
                             p_prev,

test/distributed/fsdp/test_fsdp_state_dict.py

@@ -587,7 +587,9 @@ class TestFSDPStateDict(FSDPTest):
                     model, cpu_offload.offload_params, fp16
                 )
 
-            ignore_keys = [k for k in fsdp_state_dict if NON_ROOT_FSDP_PREFIX in k]
+            ignore_keys = [
+                k for k in fsdp_state_dict.keys() if NON_ROOT_FSDP_PREFIX in k
+            ]
 
             self._validate_state_dict_contents(
                 model,
@@ -908,7 +910,7 @@ class TestFSDPStateDict(FSDPTest):
         with sd_mgr:
             fsdp_state_dict = model.state_dict()
 
-        ignore_keys = [k for k in fsdp_state_dict if NON_ROOT_FSDP_PREFIX in k]
+        ignore_keys = [k for k in fsdp_state_dict.keys() if NON_ROOT_FSDP_PREFIX in k]
         self._validate_state_dict_contents(
             model,
             fsdp_state_dict,
@@ -957,7 +959,9 @@ class TestFSDPStateDict(FSDPTest):
                 # Full name of linear_skip param tensors in SkipModel, as would be
                 # stored in checkpoint.
                 linear_skip_tensor_names = [
-                    k for k in dict(module.named_parameters()) if LINEAR_SKIP in k
+                    k
+                    for k in dict(module.named_parameters()).keys()
+                    if LINEAR_SKIP in k
                 ]
                 # skip SkipModule
                 linear_skip = getattr(module, LINEAR_SKIP)

test/distributed/launcher/api_test.py

@@ -137,7 +137,7 @@ class ElasticLaunchTest(unittest.TestCase):
         self.test_dir = tempfile.mkdtemp()
 
         # remove any lingering environment variables.
-        for env in os.environ.keys():  # noqa: SIM118
+        for env in os.environ.keys():
             if env.startswith("PET_"):
                 del os.environ[env]
 

test/distributed/launcher/test_run.py

@@ -69,7 +69,7 @@ class ElasticLaunchTest(TestCase):
         self.test_dir = tempfile.mkdtemp()
 
         # remove any lingering environment variables
-        for env in os.environ.keys():  # noqa: SIM118
+        for env in os.environ.keys():
             if env.startswith("PET_"):
                 del os.environ[env]
 

test/distributed/pipelining/test_schedule_multiproc.py

@@ -39,7 +39,6 @@ from torch.nn.modules.loss import MSELoss
 from torch.testing._internal.common_distributed import (
     MultiProcContinuousTest,
     requires_accelerator_dist_backend,
-    skip_if_lt_x_gpu,
 )
 from torch.testing._internal.common_utils import (
     check_leaked_tensors,
@@ -47,7 +46,6 @@ from torch.testing._internal.common_utils import (
     parametrize,
     run_tests,
     skip_but_pass_in_sandcastle_if,
-    TEST_MULTIACCELERATOR,
 )
 
 
@@ -58,6 +56,7 @@ batch_size = 64
 torch.manual_seed(0)
 device_type = acc.type if (acc := torch.accelerator.current_accelerator()) else "cpu"
 backend = dist.get_default_backend_for_device(device_type)
+TEST_MULTIACCELERATOR = torch.accelerator.device_count() >= 2
 
 
 @dataclass
@@ -232,7 +231,6 @@ class ScheduleTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [_ScheduleForwardOnly])
-    @skip_if_lt_x_gpu(4)
     def test_forward_only(self, ScheduleClass):
         mod, mod_ref, x, _, _ = setup_models_and_data(self.config)
         x_clone = x.clone()
@@ -276,7 +274,6 @@ class ScheduleTest(MultiProcContinuousTest):
             ScheduleInterleavedZeroBubble,
         ],
     )
-    @skip_if_lt_x_gpu(4)
     def test_eval_inference_mode(self, ScheduleClass):
         num_microbatches = 4
         if ScheduleClass in [
@@ -354,7 +351,6 @@ class ScheduleTest(MultiProcContinuousTest):
             ScheduleInterleavedZeroBubble,
         ],
     )
-    @skip_if_lt_x_gpu(4)
     def test_return_output(self, ScheduleClass):
         num_microbatches = 4
         if ScheduleClass in [
@@ -410,7 +406,6 @@ class ScheduleTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleGPipe, Schedule1F1B])
-    @skip_if_lt_x_gpu(4)
     def test_multi_iter(self, ScheduleClass):
         mod, _, x, target, loss_fn = setup_models_and_data(self.config)
         chunks = 4
@@ -434,7 +429,6 @@ class ScheduleTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleGPipe, Schedule1F1B])
-    @skip_if_lt_x_gpu(4)
     def test_kwargs_with_tracer(self, ScheduleClass):
         mod = ModelWithKwargs(d_hid, splits=self.world_size)
         mod.to(self.device)
@@ -487,7 +481,6 @@ class ScheduleTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleGPipe, Schedule1F1B])
-    @skip_if_lt_x_gpu(4)
     def test_grad_with_tracer(self, ScheduleClass):
         mod, ref_mod, x, target, loss_fn = setup_models_and_data(self.config)
 
@@ -530,7 +523,6 @@ class ScheduleTest(MultiProcContinuousTest):
     )
     @parametrize("ScheduleClass", [ScheduleGPipe, Schedule1F1B])
     @parametrize("shape_inference", [True, False])
-    @skip_if_lt_x_gpu(4)
     def test_grad_with_manual(self, ScheduleClass, shape_inference):
         mod, ref_mod, x, target, loss_fn = setup_models_and_data(self.config)
 
@@ -594,7 +586,6 @@ class ScheduleTest(MultiProcContinuousTest):
             ScheduleInterleavedZeroBubble,
         ],
     )
-    @skip_if_lt_x_gpu(4)
     def test_grad_with_manual_interleaved(self, ScheduleClass):
         stages_per_rank = 2
         n_stages = stages_per_rank * self.world_size
@@ -659,7 +650,6 @@ class ScheduleTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleInterleavedZeroBubble])
-    @skip_if_lt_x_gpu(4)
     def test_schedule_with_weight_update_mlp_e2e(self, ScheduleClass):
         stages_per_rank = 2
         n_stages = stages_per_rank * self.world_size
@@ -746,7 +736,6 @@ class ScheduleTest(MultiProcContinuousTest):
         "schedule_class",
         [ScheduleZBVZeroBubble, ScheduleDualPipeV],
     )
-    @skip_if_lt_x_gpu(4)
     def test_v_shape_schedules(self, schedule_class):
         n_stages = 8
         rank_stages = {0: [0, 7], 1: [1, 6], 2: [2, 5], 3: [3, 4]}
@@ -791,7 +780,6 @@ class ScheduleTest(MultiProcContinuousTest):
     @skip_but_pass_in_sandcastle_if(
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
-    @skip_if_lt_x_gpu(4)
     def test_custom_function_callback(self):
         """Test the custom function callback functionality with _PipelineScheduleRuntime."""
         n_stages = 8
@@ -991,7 +979,6 @@ class ScheduleTest(MultiProcContinuousTest):
         "ScheduleClass",
         [ScheduleInterleavedZeroBubble, ScheduleInterleaved1F1B],
     )
-    @skip_if_lt_x_gpu(4)
     def test_zero_bubble_with_model_kwargs(self, ScheduleClass):
         stages_per_rank = 2
         n_stages = stages_per_rank * self.world_size
@@ -1085,7 +1072,6 @@ class CustomSchedulesTest(MultiProcContinuousTest):
         "schedule_class",
         [ScheduleVShaped, ScheduleUnbalanced],
     )
-    @skip_if_lt_x_gpu(4)
     def test_non_symmetric_stage_ids(self, schedule_class):
         n_stages = schedule_class.n_stages
         rank_stages = schedule_class.rank_stages
@@ -1135,7 +1121,6 @@ class CustomSchedulesTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleWithReorderedB])
-    @skip_if_lt_x_gpu(4)
     def test_pipeline_schedule_runtime_custom_sched(self, ScheduleClass):
         n_stages = 2
         stages_per_rank = 1
@@ -1196,7 +1181,6 @@ class CustomSchedulesTest(MultiProcContinuousTest):
         not TEST_MULTIACCELERATOR, f"{backend} test requires 2+ GPUs"
     )
     @parametrize("ScheduleClass", [ScheduleWithW])
-    @skip_if_lt_x_gpu(4)
     def test_schedule_with_native_zero_bubble(self, ScheduleClass):
         n_stages = ScheduleClass.n_stages
         num_microbatches = ScheduleClass.num_microbatches

test/distributed/pipelining/test_stage.py

@@ -24,7 +24,6 @@ from torch.testing._internal.common_utils import (
     parametrize,
     run_tests,
     skip_but_pass_in_sandcastle_if,
-    TEST_MULTIACCELERATOR,
 )
 from torch.utils._pytree import tree_map_only
 
@@ -35,6 +34,7 @@ chunks = 8
 
 device_type = acc.type if (acc := torch.accelerator.current_accelerator()) else "cpu"
 backend = dist.get_default_backend_for_device(device_type)
+TEST_MULTIACCELERATOR = torch.accelerator.device_count() >= 2
 
 torch.manual_seed(0)