Merge remote-tracking branch 'origin/main' into ci_test_a_branch_that…

…_fails_to_build
NVIDIA · Aug 9, 2024 · 5c14e7c · 5c14e7c
2 parents 3b5a25a + c3f8037
commit 5c14e7c
Show file tree

Hide file tree

Showing 607 changed files with 122,350 additions and 37,026 deletions.
diff --git a/.github/workflows/blossom-ci.yml b/.github/workflows/blossom-ci.yml
@@ -24,8 +24,8 @@ jobs:
 
     # This job only runs for pull request comments
     if: |
-         contains( 'xwang233,jjsjann123,', format('{0},', github.actor)) &&
-         github.event.comment.body == '!dontusethis'
+         github.event.comment.body == '!dontusethis' &&
+         (github.actor == 'xwang233' || github.actor == 'jjsjann123')
     steps:
       - name: Check if comment is issued by authorized person
         run: blossom-ci

diff --git a/.github/workflows/lint.yml b/.github/workflows/lint.yml
@@ -14,6 +14,19 @@ env:
   working_directory: .
 
 jobs:
+  check-license:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - name: Look for missing license headers
+        working-directory: ${{ env.working_directory }}
+        run: |
+          find . -regextype posix-extended -regex '.*\.(cpp|h|py|md|svg)' | egrep -v '^./(\.|third_party|bin|build|nvfuser/include|nvfuser/version.py|tools/linter/adapters|csrc/serde/fusion_cache_generated.h)' | xargs grep -L SPDX-FileCopyrightText | tee missing-header-files.txt
+          # test that file is empty
+          test ! -s missing-header-files.txt
+
   clang-tidy:
     runs-on: ubuntu-latest
     steps:

diff --git a/.github/workflows/nvfuser-ci-trigger.yml b/.github/workflows/nvfuser-ci-trigger.yml
@@ -16,8 +16,8 @@ jobs:
 
     # This job only runs for pull request comments
     if: |
-         contains(',xwang233,jjsjann123,chang-l,csarofeen,drzejan2,IvanYashchuk,jacobhinkle,kevinstephano,liqiangxl,mmigdal-nv,naoyam,ptrblck,rdspring1,samnordmann,zasdfgbnm,crcrpar,nWEIdia,Priya2698,wujingyue,tfogal,protonu,cowanmeg,', format(',{0},', github.actor)) &&
-         startsWith(github.event.comment.body, '!build')
+         startsWith(github.event.comment.body, '!build') &&
+         (github.actor == 'xwang233' || github.actor == 'jjsjann123' || github.actor == 'chang-l' || github.actor == 'csarofeen' || github.actor == 'drzejan2' || github.actor == 'IvanYashchuk' || github.actor == 'jacobhinkle' || github.actor == 'kevinstephano' || github.actor == 'liqiangxl' || github.actor == 'mmigdal-nv' || github.actor == 'naoyam' || github.actor == 'ptrblck' || github.actor == 'rdspring1' || github.actor == 'samnordmann' || github.actor == 'zasdfgbnm' || github.actor == 'crcrpar' || github.actor == 'nWEIdia' || github.actor == 'Priya2698' || github.actor == 'wujingyue' || github.actor == 'tfogal' || github.actor == 'protonu' || github.actor == 'cowanmeg' || github.actor == 'nsarka')
     steps:
       - name: Check if comment is issued by authorized person
         run: blossom-ci

diff --git a/.gitignore b/.gitignore
@@ -48,3 +48,6 @@ foo.bin
 .pre-commit-config.yaml
 
 *_generated.*
+
+# Mac OS internal file
+.DS_Store
diff --git a/.lintrunner.toml b/.lintrunner.toml
@@ -44,7 +44,7 @@ init_command = [
     'python3',
     'tools/linter/adapters/pip_init.py',
     '--dry-run={{DRYRUN}}',
-    'clang-format==16.0.6',
+    'clang-format==18.1.8',
 ]
 is_formatter = true
 

diff --git a/CMakeLists.txt b/CMakeLists.txt
diff --git a/README.md b/README.md
@@ -1,3 +1,9 @@
+<!--
+ * SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
+ * All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+-->
+
 # Fuser
 
 A Fusion Code Generator for NVIDIA GPUs (commonly known as "nvFuser")
@@ -8,31 +14,47 @@ We publish nightly wheel packages on https://pypi.nvidia.com
 
 built-env | cuda 11.8 | cuda 12.1
 :---: | :---: | :---:
-torch 2.1 | nvfuser-cu118-torch21 | nvfuser-cu121-torch21
+torch 2.2 | nvfuser-cu118-torch22 | nvfuser-cu121-torch22
 torch nightly wheel | nvfuser-cu118 | nvfuser-cu121
 
-Note that nvfuser built against torch-2.1 isn't compatible with nightly pytorch wheel, so ensure you pick the right version suiting your environment.
+Note that nvfuser built against torch-2.2 isn't compatible with nightly pytorch wheel, so ensure you pick the right version suiting your environment.
+
+### nightly nvfuser pip wheel
 
-You can instll a given nvfuser version with `pip install --pre nvfuser-cu121 --extra-index-url https://pypi.nvidia.com`
+You can instll a nightly nvfuser pip package built against torch nightly code base with `pip install --pre nvfuser-cu121 --extra-index-url https://pypi.nvidia.com`
 
 As we build against nightly torch wheel and there's no compatibility promised on nightly wheels, we have explicitly marked the nightly torch wheel as an optinoal dependency. You can choose to install the torch wheel along with nvfuser package. e.g.
 `pip install --pre "nvfuser-cu121[torch]" --extra-index-url https://pypi.nvidia.com`.
 Note that this may uninstall your local pytorch installation and install the compatible nightly pytorch.
 
-Versioned nvfuser will be published on pypi.org [WIP]
+### nvfuser pip wheel against pytorch stable release
 
-PyPI: [https://pypi.org/project/nvfuser/](https://pypi.org/search/?q=nvfuser)
+nvfuser pip wheel built against stable torch releases is published on pypi.org. Pick the right cuda toolkit version to match your torch installation. e.g. `pip install nvfuser-cu121-torch22`
 
+PyPI: [https://pypi.org/project/nvfuser/](https://pypi.org/search/?q=nvfuser)
 
 ## Developer
 
 Docs: https://github.com/NVIDIA/Fuser/wiki
 
 Supported compilers:
-- gcc 11.4+
-- clang14+
+
+**GCC:**
+
+We support all "supported releases" of gcc as specified in [the official site](https://gcc.gnu.org/).
+As of 6/20/2024, they are:
+
+- gcc 11.4
+- gcc 12.4
+- gcc 13.2
+- gcc 14.1
+
+**Clang:**
+
+- clang 14+
 
 Supported C++ standard:
+
 - C++17
 - C++20
 

diff --git a/benchmark/batch_norm_channels_first.cpp → benchmarks/cpp/batch_norm_channels_first.cpp b/benchmark/batch_norm_channels_first.cpp → benchmarks/cpp/batch_norm_channels_first.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/...rk/batch_norm_channels_first_backward.cpp → ...pp/batch_norm_channels_first_backward.cpp b/...rk/batch_norm_channels_first_backward.cpp → ...pp/batch_norm_channels_first_backward.cpp
@@ -21,8 +21,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/batch_norm_channels_last.cpp → benchmarks/cpp/batch_norm_channels_last.cpp b/benchmark/batch_norm_channels_last.cpp → benchmarks/cpp/batch_norm_channels_last.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/...ark/batch_norm_channels_last_backward.cpp → ...cpp/batch_norm_channels_last_backward.cpp b/...ark/batch_norm_channels_last_backward.cpp → ...cpp/batch_norm_channels_last_backward.cpp
@@ -21,8 +21,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/bert.cpp → benchmarks/cpp/bert.cpp b/benchmark/bert.cpp → benchmarks/cpp/bert.cpp
@@ -21,8 +21,8 @@
 
 #include <sstream>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/broadcast.cpp → benchmarks/cpp/broadcast.cpp b/benchmark/broadcast.cpp → benchmarks/cpp/broadcast.cpp
@@ -19,8 +19,8 @@
 
 #include <sstream>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/gelu_backward.cpp → benchmarks/cpp/gelu_backward.cpp b/benchmark/gelu_backward.cpp → benchmarks/cpp/gelu_backward.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/gelu_backward_reduction.cpp → benchmarks/cpp/gelu_backward_reduction.cpp b/benchmark/gelu_backward_reduction.cpp → benchmarks/cpp/gelu_backward_reduction.cpp
@@ -19,8 +19,8 @@
 
 #include <sstream>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 
@@ -114,9 +114,12 @@ static void NvFuserScheduler_GeluBackwardReduction(
 
   runBenchmarkIterations(benchmark_state, fusion_executor_cache, aten_inputs);
 
+  // inputs: gradient tensor + input tensor
+  // outputs: output, output_of_reduction
   benchmark_state.SetBytesProcessed(
       int64_t(benchmark_state.iterations()) *
-      (iter_size * reduction_size + iter_size) * int64_t(dataTypeSize(dtype)));
+      (iter_size * reduction_size * 3 + iter_size) *
+      int64_t(dataTypeSize(dtype)));
 }
 
 static void Baseline_GeluBackwardReduction(
@@ -172,7 +175,8 @@ static void Baseline_GeluBackwardReduction(
 
   benchmark_state.SetBytesProcessed(
       int64_t(benchmark_state.iterations()) *
-      (iter_size * reduction_size + iter_size) * int64_t(dataTypeSize(dtype)));
+      (iter_size * reduction_size * 3 + iter_size) *
+      int64_t(dataTypeSize(dtype)));
 }
 
 //------------------------------------------------------------------------------

diff --git a/benchmark/heuristic_cache.cpp → benchmarks/cpp/heuristic_cache.cpp b/benchmark/heuristic_cache.cpp → benchmarks/cpp/heuristic_cache.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/heuristic_lookup.cpp → benchmarks/cpp/heuristic_lookup.cpp b/benchmark/heuristic_lookup.cpp → benchmarks/cpp/heuristic_lookup.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/indexselect.cpp → benchmarks/cpp/indexselect.cpp b/benchmark/indexselect.cpp → benchmarks/cpp/indexselect.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/instance_norm.cpp → benchmarks/cpp/instance_norm.cpp b/benchmark/instance_norm.cpp → benchmarks/cpp/instance_norm.cpp
@@ -17,8 +17,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/layer_norm.cpp → benchmarks/cpp/layer_norm.cpp b/benchmark/layer_norm.cpp → benchmarks/cpp/layer_norm.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/layer_norm_backward.cpp → benchmarks/cpp/layer_norm_backward.cpp b/benchmark/layer_norm_backward.cpp → benchmarks/cpp/layer_norm_backward.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/layer_norm_fused.cpp → benchmarks/cpp/layer_norm_fused.cpp b/benchmark/layer_norm_fused.cpp → benchmarks/cpp/layer_norm_fused.cpp
@@ -19,8 +19,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 
@@ -59,7 +59,7 @@ static void setupLayerNormFused(Fusion* fusion, DataType dtype) {
   auto tv19 = broadcast(tv18, {false, true});
 
   nvfuser::Val* num_features = IrBuilder::create<Val>(1.0);
-  num_features = mul(num_features, tv0->getLeafDomain()[0]->extent());
+  num_features = mul(num_features, tv0->getLogicalDomain()[0]->extent());
   auto s20 = num_features;
 
   auto s21 = reciprocal(s20);

diff --git a/benchmark/lstm_cell.cpp → benchmarks/cpp/lstm_cell.cpp b/benchmark/lstm_cell.cpp → benchmarks/cpp/lstm_cell.cpp
@@ -15,8 +15,8 @@
 
 #include <cuda_runtime.h>
 
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 

diff --git a/benchmark/main.cpp → benchmarks/cpp/main.cpp b/benchmark/main.cpp → benchmarks/cpp/main.cpp
diff --git a/benchmark/many_pointwise_ops.cpp → benchmarks/cpp/many_pointwise_ops.cpp b/benchmark/many_pointwise_ops.cpp → benchmarks/cpp/many_pointwise_ops.cpp
@@ -11,8 +11,8 @@
 #include <ops/arith.h>
 
 #include <benchmark/benchmark.h>
-#include <benchmark/utils.h>
-#include <test/utils.h>
+#include <benchmarks/cpp/utils.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;