2022-11-05 nightly release (957a9b6)

pytorch · ezyang · Sep 11, 2022 · Sep 12, 2022 · Sep 13, 2022 · Sep 14, 2022
commit 12d78465ebfc07a72f924c67dde7f5cb1c439667
@@ -52,7 +52,8 @@ runs:
         .github/scripts/filter_test_configs.py \
           --test-matrix "${{ inputs.test-matrix }}" \
           --pr-number "${{ github.event.pull_request.number }}" \
-          --tag "${{ steps.parse-ref.outputs.tag }}"
+          --tag "${{ steps.parse-ref.outputs.tag }}" \
+          --event-name "${{ github.event_name }}"
 
     - name: Print the filtered test matrix
       shell: bash

@@ -42,26 +42,26 @@ def should_check(filename: Path) -> bool:
             print("ERROR: duplicate workflow name:", name, file=sys.stderr)
             errors_found = True
         names.add(name)
-
-        expected = {
-            "group": EXPECTED_GROUP,
-            "cancel-in-progress": True,
-        }
-        actual = data.get("concurrency", None)
-        if actual != expected:
+        actual = data.get("concurrency", {})
+        if not actual.get("group", "").startswith(EXPECTED_GROUP):
             print(
                 f"'concurrency' incorrect or not found in '{filename.relative_to(REPO_ROOT)}'",
                 file=sys.stderr,
             )
             print(
-                f"expected: {expected}",
+                f"concurrency group should start with {EXPECTED_GROUP} but found {actual.get('group', None)}",
                 file=sys.stderr,
             )
+            errors_found = True
+        if not actual.get("cancel-in-progress", False):
             print(
-                f"actual:   {actual}",
+                f"'concurrency' incorrect or not found in '{filename.relative_to(REPO_ROOT)}'",
+                file=sys.stderr,
+            )
+            print(
+                f"concurrency cancel-in-progress should be True but found {actual.get('cancel-in-progress', None)}",
                 file=sys.stderr,
             )
-            errors_found = True
 
     if errors_found:
         sys.exit(1)
@@ -38,6 +38,7 @@ def parse_args() -> Any:
     parser.add_argument("--test-matrix", type=str, required=True, help="the original test matrix")
     parser.add_argument("--pr-number", type=str, help="the pull request number")
     parser.add_argument("--tag", type=str, help="the associated tag if it exists")
+    parser.add_argument("--event-name", type=str, help="name of the event that triggered the job (pull, schedule, etc)")
     return parser.parse_args()
 
 
@@ -159,6 +160,10 @@ def main() -> None:
         # No PR number, no tag, we can just return the test matrix as it is
         filtered_test_matrix = test_matrix
 
+    if args.event_name == "schedule":
+        for config in filtered_test_matrix.get("include", []):
+            config["mem_leak_check"] = "mem_leak_check"
+
     # Set the filtered test matrix as the output
     set_output("test-matrix", json.dumps(filtered_test_matrix))
 

@@ -122,6 +122,7 @@ jobs:
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
           XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
+          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0'}}
         timeout-minutes: 240
         run: |
           set -x
@@ -175,6 +176,7 @@ jobs:
             -e SCCACHE_S3_KEY_PREFIX \
             -e XLA_CUDA \
             -e XLA_CLANG_CACHE_S3_BUCKET_NAME \
+            -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --security-opt seccomp=unconfined \

@@ -128,6 +128,8 @@ jobs:
 
       - name: Test
         id: test
+        env:
+          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0'}}
         run: |
           COMMIT_MESSAGES=$(git cherry -v "origin/${GIT_DEFAULT_BRANCH:-master}")
 

@@ -97,6 +97,7 @@ jobs:
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           XLA_CLANG_CACHE_S3_BUCKET_NAME: ossci-compiler-clang-cache-circleci-xla
           PYTORCH_JIT_ENABLE_NVFUSER: 1
+          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0'}}
         timeout-minutes: 270
         run: |
           set -x
@@ -146,6 +147,7 @@ jobs:
             -e MAX_JOBS="$(nproc --ignore=2)" \
             -e SCCACHE_BUCKET \
             -e XLA_CLANG_CACHE_S3_BUCKET_NAME \
+            -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --security-opt seccomp=unconfined \

@@ -124,6 +124,7 @@ jobs:
           TEST_CONFIG: ${{ matrix.config }}
           PR_BODY: ${{ github.event.pull_request.body }}
           TORCH_CUDA_ARCH_LIST: "7.0"
+          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0'}}
         run: |
           COMMIT_MESSAGES=$(git cherry -v "origin/${GIT_DEFAULT_BRANCH:-master}")
 

@@ -1,9 +1,12 @@
 name: inductor
 
 on:
+  schedule:
+    - cron: 45 1,5,9,13,17,21 * * *
   push:
     tags:
       - ciflow/inductor/*
+      - ciflow/periodic/*
   workflow_dispatch:
 
 concurrency:

@@ -9,7 +9,7 @@ on:
   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' }}
+  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' }}
   cancel-in-progress: true
 
 jobs:

@@ -9,9 +9,11 @@ on:
       - release/*
       - landchecks/*
   workflow_dispatch:
+  schedule:
+    - cron: 29 8 * * *  # about 1:29am PDT
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
   cancel-in-progress: true
 
 jobs:

@@ -10,9 +10,11 @@ on:
     tags:
       - ciflow/trunk/*
   workflow_dispatch:
+  schedule:
+    - cron: 29 8 * * *  # about 1:29am PDT
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  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' }}
   cancel-in-progress: true
 
 jobs:

diff --git a/.jenkins/pytorch/test.sh b/.jenkins/pytorch/test.sh
@@ -109,14 +109,6 @@ if [[ "$TEST_CONFIG" == *inductor* ]]; then
   export PYTORCH_TEST_WITH_INDUCTOR=1
 fi
 
-# TODO: this condition is never true, need to fix this.
-if [[ -n "$PR_NUMBER" ]] && [[ -z "$CI_MASTER" || "$CI_MASTER" == "false" ]]; then
-  # skip expensive checks when on PR and CI_MASTER flag is not set
-  export PYTORCH_TEST_SKIP_CUDA_MEM_LEAK_CHECK=1
-else
-  export PYTORCH_TEST_SKIP_CUDA_MEM_LEAK_CHECK=0
-fi
-
 if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
   # Print GPU info
   rocminfo

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -285,6 +285,7 @@ if(NOT USE_XNNPACK AND CMAKE_VERSION VERSION_LESS ${XNNPACK_MIN_CMAKE_VER})
 endif()
 option(USE_ZMQ "Use ZMQ" OFF)
 option(USE_ZSTD "Use ZSTD" OFF)
+option(TORCH_DISABLE_GPU_ASSERTS "Disable GPU asserts by default" OFF)
 # Ensure that an ITT build is the default for x86 CPUs
 cmake_dependent_option(
   USE_ITT "Use Intel(R) VTune Profiler ITT functionality" ON

diff --git a/aten/src/ATen/core/interned_strings.h b/aten/src/ATen/core/interned_strings.h
@@ -50,8 +50,11 @@ namespace c10 {
   _(prim, FunctionalGraph)           \
   _(prim, add_optional)              \
   _(prim, view_copy)                 \
+  _(prim, permute_copy)              \
   _(prim, reshape_copy)              \
   _(prim, squeeze_copy)              \
+  _(prim, t_copy)                    \
+  _(prim, transpose_copy)            \
   _(prim, unsqueeze_copy)            \
   _(prim, flatten_copy)              \
   _(prim, expand_copy)               \

diff --git a/aten/src/ATen/native/README.md b/aten/src/ATen/native/README.md
@@ -47,10 +47,9 @@ signature.
   if one argument is a `FloatTensor`, all other arguments are checked
   to be `FloatTensor`s).
   `Tensor` or `Tensor?` must sometimes be annotated to indicate aliasing and mutability.
-  In general annotations can be defined via the following four situations:
-  - `Tensor(a)` - `a` is a set of Tensors that may alias to the same data.
+  In general annotations can be defined via the following situations:
+  - `Tensor(a)` - `a` is a set of Tensors that may alias to the same data. The set could have a size of one.
   - `Tensor(a!)` - members of `a` may be written to thus mutating the underlying data.
-  - `Tensor!` - shorthand for Tensor(fresh\_identifier!)
   - `Tensor(a! -> a|b)` - Tensor is in set `a`, written to, and after the write is in set `a` AND `b`.
   For more details on when and why this needs to happen, please see the section on annotations.
 - `Tensor[]`.  A `Tensor[]` argument translates into a C++ argument of type `ArrayRef<Tensor>`

diff --git a/aten/src/ATen/native/TensorAdvancedIndexing.cpp b/aten/src/ATen/native/TensorAdvancedIndexing.cpp
@@ -1512,6 +1512,10 @@ TORCH_IMPL_FUNC(scatter_src_out)
  const Tensor& index,
  const Tensor& src,
  const Tensor& out) {
+  // See note [Writing Nondeterministic Operations]
+  // Nondeterministic when index contains duplicate entries, src is a tensor,
+  // and reduce=None
+  at::globalContext().alertNotDeterministic("scatter with src tensor and reduce=None");
   scatter_impl(self, dim, index, src, out,
                scatter_reduce_stub,
                scatter_stub);

@@ -645,11 +645,14 @@ Tensor embedding_dense_backward_mps(
 
             MPSGraphTensor* indicesTensor = native_mps::mpsGraphUnrankedPlaceHolder(mpsGraph, native_mps::getMPSDataType(indices.scalar_type()));
 
-            MPSGraphTensor *reshapedIndicesTensor = [mpsGraph  expandDimsOfTensor:indicesTensor
-                             axes:@[@-1]
-                             name:nil];
+            MPSGraphTensor* reshapedIndicesTensor = indicesTensor;
 
-            MPSGraphTensor *outgoingGradTensor;
+            if (num_indices_dims != 0)
+              reshapedIndicesTensor = [mpsGraph  expandDimsOfTensor:indicesTensor
+                              axes:@[@-1]
+                              name:nil];
+
+            MPSGraphTensor* outgoingGradTensor;
             outgoingGradTensor = [mpsGraph scatterNDWithUpdatesTensor:incomingGradTensor
                             indicesTensor:reshapedIndicesTensor
                                     shape:native_mps::getMPSShape(IntArrayRef(outgoing_gradient_shape.data(), outgoing_gradient_shape.size()))

diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
@@ -2485,11 +2485,17 @@
   device_check: NoCheck   # TensorIterator
   structured_delegate: frac.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: frac_sparse
+    SparseCsrCPU, SparseCsrCUDA: frac_sparse_csr
 
 - func: frac_(Tensor(a!) self) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
   structured_delegate: frac.out
   variants: function, method
+  dispatch:
+    SparseCPU, SparseCUDA: frac_sparse_
+    SparseCsrCPU, SparseCsrCUDA: frac_sparse_csr_
 
 - func: frac.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
@@ -2498,6 +2504,8 @@
   dispatch:
     CPU, CUDA: frac_out
     MPS: frac_out_mps
+    SparseCPU, SparseCUDA: frac_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: frac_sparse_csr_out
 
 - func: full.names(int[] size, Scalar fill_value, *, Dimname[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   device_check: NoCheck
@@ -4026,15 +4034,21 @@
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: deg2rad
+    SparseCPU, SparseCUDA: deg2rad_sparse
+    SparseCsrCPU, SparseCsrCUDA: deg2rad_sparse_csr
 
 - func: deg2rad_(Tensor(a!) self) -> Tensor(a!)
   variants: function, method
   dispatch:
     CompositeExplicitAutograd: deg2rad_
+    SparseCPU, SparseCUDA: deg2rad_sparse_
+    SparseCsrCPU, SparseCsrCUDA: deg2rad_sparse_csr_
 
 - func: deg2rad.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
   dispatch:
     CompositeExplicitAutograd: deg2rad_out
+    SparseCPU, SparseCUDA: deg2rad_sparse_out
+    SparseCsrCPU, SparseCsrCUDA: deg2rad_sparse_csr_out
 
 - func: scalar_tensor(Scalar s, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
   dispatch:

@@ -327,14 +327,15 @@ void pytorch_q8gemm_ukernel_4x4c2__sse2(
         (uint32_t)_mm_cvtsi128_si32(_mm_unpackhi_epi32(vout, vout));
     *((uint32_t*)c3) = (uint32_t)_mm_cvtsi128_si32(_mm_srli_si128(vout, 12));
   } else {
+    typedef PYTORCH_QNNP_UNALIGNED uint16_t unaligned_uint16_t;
     if (nr >= 2) {
-      *((uint16_t*)c0) = (uint16_t)_mm_extract_epi16(vout, 0);
+      *((unaligned_uint16_t*)c0) = (uint16_t)_mm_extract_epi16(vout, 0);
       c0 += 2;
-      *((uint16_t*)c1) = (uint16_t)_mm_extract_epi16(vout, 2);
+      *((unaligned_uint16_t*)c1) = (uint16_t)_mm_extract_epi16(vout, 2);
       c1 += 2;
-      *((uint16_t*)c2) = (uint16_t)_mm_extract_epi16(vout, 4);
+      *((unaligned_uint16_t*)c2) = (uint16_t)_mm_extract_epi16(vout, 4);
       c2 += 2;
-      *((uint16_t*)c3) = (uint16_t)_mm_extract_epi16(vout, 6);
+      *((unaligned_uint16_t*)c3) = (uint16_t)_mm_extract_epi16(vout, 6);
       c3 += 2;
       vout = _mm_srli_epi32(vout, 16);
       nr -= 2;

@@ -80,3 +80,15 @@
 #if defined(_MSC_VER)
 #define __builtin_prefetch
 #endif
+
+#if defined(__GNUC__)
+  #define PYTORCH_QNNP_UNALIGNED __attribute__((__aligned__(1)))
+#elif defined(_MSC_VER)
+  #if defined(_M_IX86)
+    #define PYTORCH_QNNP_UNALIGNED
+  #else
+    #define PYTORCH_QNNP_UNALIGNED __unaligned
+  #endif
+#else
+  #error "Platform-specific implementation of PYTORCH_QNNP_UNALIGNED required"
+#endif
diff --git a/aten/src/ATen/native/sparse/SparseCsrTensorMath.cpp b/aten/src/ATen/native/sparse/SparseCsrTensorMath.cpp
@@ -48,6 +48,8 @@
 #include <ATen/ops/conj_physical.h>
 #include <ATen/ops/conj_physical_native.h>
 #include <ATen/ops/copy_native.h>
+#include <ATen/ops/deg2rad.h>
+#include <ATen/ops/deg2rad_native.h>
 #include <ATen/ops/empty.h>
 #include <ATen/ops/erf.h>
 #include <ATen/ops/erf_native.h>
@@ -58,6 +60,8 @@
 #include <ATen/ops/fill_native.h>
 #include <ATen/ops/floor.h>
 #include <ATen/ops/floor_native.h>
+#include <ATen/ops/frac.h>
+#include <ATen/ops/frac_native.h>
 #include <ATen/ops/isinf.h>
 #include <ATen/ops/isinf_native.h>
 #include <ATen/ops/isnan.h>
@@ -355,10 +359,12 @@ CREATE_UNARY_UFUNC(asinh);
 CREATE_UNARY_UFUNC(atan);
 CREATE_UNARY_UFUNC(atanh);
 CREATE_UNARY_UFUNC(ceil);
+CREATE_UNARY_UFUNC(deg2rad);
 CREATE_UNARY_UFUNC(erf);
 CREATE_UNARY_UFUNC(erfinv);
 CREATE_UNARY_UFUNC(expm1);
 CREATE_UNARY_UFUNC(floor);
+CREATE_UNARY_UFUNC(frac);
 CREATE_UNARY_UFUNC(log1p);
 CREATE_UNARY_UFUNC(neg);
 CREATE_UNARY_UFUNC(rad2deg);

diff --git a/aten/src/ATen/native/sparse/SparseUnaryOps.cpp b/aten/src/ATen/native/sparse/SparseUnaryOps.cpp
@@ -19,6 +19,8 @@
 #include <ATen/ops/atanh_native.h>
 #include <ATen/ops/ceil.h>
 #include <ATen/ops/ceil_native.h>
+#include <ATen/ops/deg2rad.h>
+#include <ATen/ops/deg2rad_native.h>
 #include <ATen/ops/erf.h>
 #include <ATen/ops/erf_native.h>
 #include <ATen/ops/erfinv.h>
@@ -27,6 +29,8 @@
 #include <ATen/ops/expm1_native.h>
 #include <ATen/ops/floor.h>
 #include <ATen/ops/floor_native.h>
+#include <ATen/ops/frac.h>
+#include <ATen/ops/frac_native.h>
 #include <ATen/ops/isinf.h>
 #include <ATen/ops/isinf_native.h>
 #include <ATen/ops/isnan.h>
@@ -165,10 +169,12 @@ COALESCED_UNARY_UFUNC(asinh);
 COALESCED_UNARY_UFUNC(atan);
 COALESCED_UNARY_UFUNC(atanh);
 COALESCED_UNARY_UFUNC(ceil);
+COALESCED_UNARY_UFUNC(deg2rad);
 COALESCED_UNARY_UFUNC(erf);
 COALESCED_UNARY_UFUNC(erfinv);
 COALESCED_UNARY_UFUNC(expm1);
 COALESCED_UNARY_UFUNC(floor);
+COALESCED_UNARY_UFUNC(frac);
 COALESCED_UNARY_UFUNC(log1p);
 COALESCED_UNARY_UFUNC(round);
 COALESCED_UNARY_UFUNC(sign);