2023-04-25 nightly release (ecd2c71)

pytorch · ezyang · Sep 11, 2022 · Sep 12, 2022 · Sep 13, 2022 · Sep 14, 2022
commit 015e4ddb7243e324cc81f0c02566edf48eab4a09
@@ -47,11 +47,7 @@ compile_arm64() {
 }
 
 compile_x86_64() {
-  USE_DISTRIBUTED=0 WERROR=1 python setup.py bdist_wheel
-}
-
-compile_x86_64() {
-  USE_DISTRIBUTED=0 WERROR=1 python setup.py bdist_wheel
+  USE_DISTRIBUTED=0 WERROR=1 python setup.py bdist_wheel --plat-name=macosx_10_9_x86_64
 }
 
 build_lite_interpreter() {

@@ -9,7 +9,7 @@ sysctl -a | grep machdep.cpu
 
 # These are required for both the build job and the test job.
 # In the latter to test cpp extensions.
-export MACOSX_DEPLOYMENT_TARGET=10.9
+export MACOSX_DEPLOYMENT_TARGET=11.0
 export CXX=clang++
 export CC=clang
 

diff --git a/.flake8 b/.flake8
@@ -1,4 +1,6 @@
 [flake8]
+# NOTE: **Mirror any changes** to this file the [tool.ruff] config in pyproject.toml
+# before we can fully move to use ruff
 enable-extensions = G
 select = B,C,E,F,G,P,SIM1,T4,W,B9
 max-line-length = 120

@@ -53,9 +53,9 @@ body:
       description: |
         Please run the following and paste the output below.
         ```sh
-        wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
+        curl -OL https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
         # For security purposes, please check the contents of collect_env.py before running it.
-        python collect_env.py
+        python3 collect_env.py
         ```
     validations:
       required: true
@@ -1 +1 @@
-f235d4da06905b35d75879a0a9bc3034ab7385ac
+486b32e5b1a2e2964a7db4c7db25b59d879b8796
@@ -420,6 +420,7 @@
   - gchanan
   - ezyang
   - dzhulgakov
+  - malfet
   mandatory_checks_name:
   - EasyCLA
   - Lint

diff --git a/.lintrunner.toml b/.lintrunner.toml
@@ -972,3 +972,29 @@ command = [
     'python3',
     'tools/linter/adapters/lintrunner_version_linter.py'
 ]
+
+[[linter]]
+code = 'RUFF'
+include_patterns = ['**/*.py']
+exclude_patterns = [
+    'caffe2/**',
+    'functorch/docs/**',
+    'functorch/notebooks/**',
+    'scripts/**',
+    'third_party/**',
+]
+command = [
+    'python3',
+    'tools/linter/adapters/ruff_linter.py',
+    '--config=pyproject.toml',
+    '--show-disable',
+    '--',
+    '@{{PATHSFILE}}'
+]
+init_command = [
+    'python3',
+    'tools/linter/adapters/pip_init.py',
+    '--dry-run={{DRYRUN}}',
+    'ruff==0.0.262',
+]
+is_formatter = true
diff --git a/README.md b/README.md
@@ -439,7 +439,7 @@ To learn more about making a contribution to Pytorch, please see our [Contributi
 
 PyTorch is a community-driven project with several skillful engineers and researchers contributing to it.
 
-PyTorch is currently maintained by [Adam Paszke](https://apaszke.github.io/), [Sam Gross](https://github.com/colesbury), [Soumith Chintala](http://soumith.ch) and [Gregory Chanan](https://github.com/gchanan) with major contributions coming from hundreds of talented individuals in various forms and means.
+PyTorch is currently maintained by [Soumith Chintala](http://soumith.ch), [Gregory Chanan](https://github.com/gchanan), [Dmytro Dzhulgakov](https://github.com/dzhulgakov), [Edward Yang](https://github.com/ezyang), and [Nikita Shulga](https://github.com/malfet) with major contributions coming from hundreds of talented individuals in various forms and means.
 A non-exhaustive but growing list needs to mention: Trevor Killeen, Sasank Chilamkurthy, Sergey Zagoruyko, Adam Lerer, Francisco Massa, Alykhan Tejani, Luca Antiga, Alban Desmaison, Andreas Koepf, James Bradbury, Zeming Lin, Yuandong Tian, Guillaume Lample, Marat Dukhan, Natalia Gimelshein, Christian Sarofeen, Martin Raison, Edward Yang, Zachary Devito.
 
 Note: This project is unrelated to [hughperkins/pytorch](https://github.com/hughperkins/pytorch) with the same name. Hugh is a valuable contributor to the Torch community and has helped with many things Torch and PyTorch.

diff --git a/aten/src/ATen/CachedTensorUtils.cpp b/aten/src/ATen/CachedTensorUtils.cpp
@@ -0,0 +1,49 @@
+#include <ATen/ATen.h>
+
+#include <ATen/CachedTensorUtils.h>
+
+
+namespace at {
+namespace caching {
+
+
+using weakref_type = c10::weak_intrusive_ptr<TensorImpl, UndefinedTensorImpl>;
+
+bool cached_tensorimpls_enabled = false;
+
+// Like `cached_casts` in autocast_mode, we hash on the TensorImpl*
+//  and keep the pointer alive with a weakref value.
+ska::flat_hash_map<TensorImpl*, weakref_type> cached_tensorimpls;
+std::mutex cached_tensorimpl_mutex;
+
+
+bool is_cached_tensor(const at::Tensor& t) {
+  if (!cached_tensorimpls_enabled) {
+    return false;
+  }
+  const std::lock_guard<std::mutex> lock(cached_tensorimpl_mutex);
+  return cached_tensorimpls.count(t.unsafeGetTensorImpl());
+}
+
+void add_cached_tensor(const at::Tensor& t) {
+  TORCH_INTERNAL_ASSERT(cached_tensorimpls_enabled);
+  const std::lock_guard<std::mutex> lock(cached_tensorimpl_mutex);
+  cached_tensorimpls.emplace(t.unsafeGetTensorImpl(), weakref_type(t.getIntrusivePtr()));
+}
+
+void remove_cached_tensor(const at::Tensor& t) {
+  TORCH_INTERNAL_ASSERT(cached_tensorimpls_enabled);
+  const std::lock_guard<std::mutex> lock(cached_tensorimpl_mutex);
+  cached_tensorimpls.erase(t.unsafeGetTensorImpl());
+}
+
+void set_cached_tensors_enabled(bool enabled) {
+  cached_tensorimpls_enabled = enabled;
+}
+
+size_t adjusted_use_count(const at::Tensor& t) {
+  return t.use_count() - (is_cached_tensor(t) ? 1 : 0);
+}
+
+}
+}
diff --git a/aten/src/ATen/CachedTensorUtils.h b/aten/src/ATen/CachedTensorUtils.h
@@ -0,0 +1,26 @@
+#pragma once
+
+#include <ATen/ATen.h>
+
+namespace at {
+namespace caching {
+
+// Some systems (just cudagraphs currently) will persist a static tensor output
+// whose TensorImpl does not change across iterations. For these tensors caching
+// dtype conversions is invalid. Additionally, there will be an extra reference
+// count to these cached tensors that would prevent buffer inplacing and other
+// checks on tensor uniqueness. If we are not using these systems the enabled
+// flag will be false and we will avoid the hash lookup.
+
+TORCH_API bool is_cached_tensor(const at::Tensor& t);
+TORCH_API void add_cached_tensor(const at::Tensor& t);
+TORCH_API void remove_cached_tensor(const at::Tensor& t);
+TORCH_API void set_cached_tensors_enabled(bool enable);
+
+// For gradient buffer stealing we will adjust the use count of tensors
+// which are persisted by cudagraphs, just as we need to adjust reference
+// count of tensors with hooks.
+TORCH_API size_t adjusted_use_count(const at::Tensor& t);
+
+} // namespace caching
+} // namespace at
diff --git a/aten/src/ATen/autocast_mode.cpp b/aten/src/ATen/autocast_mode.cpp
@@ -3,6 +3,7 @@
 #include <iostream>
 #include <exception>
 #include <mutex>
+#include <ATen/CachedTensorUtils.h>
 
 namespace at {
 namespace autocast {
@@ -66,9 +67,10 @@ namespace {
 // directly against incoming TensorImpl*s.
 using weakref_type = c10::weak_intrusive_ptr<TensorImpl, UndefinedTensorImpl>;
 using val_type = std::tuple<weakref_type, Tensor>;
-std::unordered_map<TensorImpl*, val_type> cached_casts;
+ska::flat_hash_map<TensorImpl*, val_type> cached_casts;
 std::mutex cached_casts_mutex;
 
+
 // nesting tracks the nesting depth of the Python-side context manager.
 // When the autocast context manager exits to a nesting level that's outside
 // any instance of autocast (which should occur at the end of each forward pass)
@@ -168,7 +170,9 @@ Tensor cached_cast(at::ScalarType to_type, const Tensor& arg, DeviceType device_
     // See cached_casts declaration above for detailed strategy.
     bool can_try_cache = (to_type == get_lower_precision_fp_from_device_type(device_type) &&
                          arg.scalar_type() == at::kFloat && arg.requires_grad() &&
-                         arg.is_leaf() && !arg.is_view() && cache_enabled);
+                         arg.is_leaf() && !arg.is_view() && cache_enabled &&
+                         !at::caching::is_cached_tensor(arg));
+
     if (can_try_cache) {
       const std::lock_guard<std::mutex> lock(cached_casts_mutex);
       auto it = cached_casts.find(arg.unsafeGetTensorImpl());

diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_complex_double.h b/aten/src/ATen/cpu/vec/vec256/vec256_complex_double.h
@@ -121,7 +121,10 @@ template <> class Vectorized<c10::complex<double>> {
     return _mm256_hadd_pd(val_2, val_2);            // a*a+b*b a*a+b*b
   }
   __m256d abs_() const {
-    return _mm256_sqrt_pd(abs_2_());                // abs     abs
+    auto real = _mm256_movedup_pd(values);       // real real
+    // movehdup_pd does not exist...
+    auto imag = _mm256_permute_pd(values, 0xf);  // imag imag
+    return Sleef_hypotd4_u05(real, imag);        // abs  abs
   }
   Vectorized<c10::complex<double>> abs() const {
     const __m256d real_mask = _mm256_castsi256_pd(_mm256_setr_epi64x(0xFFFFFFFFFFFFFFFF, 0x0000000000000000,
@@ -143,11 +146,8 @@ template <> class Vectorized<c10::complex<double>> {
     auto abs = abs_();
     auto zero = _mm256_setzero_pd();
     auto mask = _mm256_cmp_pd(abs, zero, _CMP_EQ_OQ);
-    auto abs_val = Vectorized(abs);
-
-    auto div = values / abs_val.values;       // x / abs(x)
-
-    return blendv(div, zero, mask);
+    auto div = values / abs;
+    return _mm256_blendv_pd(div, zero, mask);
   }
   __m256d real_() const {
     const __m256d real_mask = _mm256_castsi256_pd(_mm256_setr_epi64x(0xFFFFFFFFFFFFFFFF, 0x0000000000000000,

diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_complex_float.h b/aten/src/ATen/cpu/vec/vec256/vec256_complex_float.h
@@ -157,7 +157,9 @@ template <> class Vectorized<c10::complex<float>> {
     return _mm256_permute_ps(ret, 0xD8);
   }
   __m256 abs_() const {
-    return _mm256_sqrt_ps(abs_2_());                // abs     abs
+    auto real = _mm256_moveldup_ps(values);   // real real
+    auto imag = _mm256_movehdup_ps(values);   // imag imag
+    return Sleef_hypotf8_u05(real, imag);     // abs  abs
   }
   Vectorized<c10::complex<float>> abs() const {
     const __m256 real_mask = _mm256_castsi256_ps(_mm256_setr_epi32(0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0x00000000,
@@ -179,10 +181,7 @@ template <> class Vectorized<c10::complex<float>> {
     auto abs = abs_();
     auto zero = _mm256_setzero_ps();
     auto mask = _mm256_cmp_ps(abs, zero, _CMP_EQ_OQ);
-    auto abs_val = Vectorized(abs);
-
-    auto div = values / abs_val.values;       // x / abs(x)
-
+    auto div = values / abs;
     return _mm256_blendv_ps(div, zero, mask);
   }
   __m256 real_() const {

diff --git a/aten/src/ATen/cpu/vec/vec512/vec512_complex_double.h b/aten/src/ATen/cpu/vec/vec512/vec512_complex_double.h
@@ -174,7 +174,10 @@ template <> class Vectorized<c10::complex<double>> {
     return hadd_pd(val_2, val_2);            // a*a+b*b a*a+b*b
   }
   __m512d abs_() const {
-    return _mm512_sqrt_pd(abs_2_());                // abs     abs
+    auto real = _mm512_movedup_pd(values);        // real real
+    // movehdup_pd does not exist...
+    auto imag = _mm512_permute_pd(values, 0xff);  // imag imag
+    return Sleef_hypotd8_u05(real, imag);         // abs  abs
   }
   Vectorized<c10::complex<double>> abs() const {
     const __m512d real_mask = _mm512_castsi512_pd(_mm512_setr_epi64(0xFFFFFFFFFFFFFFFF, 0x0000000000000000,
@@ -200,13 +203,8 @@ template <> class Vectorized<c10::complex<double>> {
     auto abs = abs_();
     auto zero = _mm512_setzero_pd();
     auto mask = _mm512_cmp_pd_mask(abs, zero, _CMP_EQ_OQ);
-    auto mask_vec = _mm512_mask_set1_epi64(_mm512_castpd_si512(zero), mask,
-                                          0xFFFFFFFFFFFFFFFF);
-    auto abs_val = Vectorized(abs);
-
-    auto div = values / abs_val.values;       // x / abs(x)
-
-    return blendv(div, zero, _mm512_castsi512_pd(mask_vec));
+    auto div = values / abs;
+    return _mm512_mask_blend_pd(mask, div, zero);
   }
   __m512d real_() const {
     const __m512d real_mask = _mm512_castsi512_pd(_mm512_setr_epi64(0xFFFFFFFFFFFFFFFF, 0x0000000000000000,

diff --git a/aten/src/ATen/cpu/vec/vec512/vec512_complex_float.h b/aten/src/ATen/cpu/vec/vec512/vec512_complex_float.h
@@ -680,7 +680,9 @@ template <> class Vectorized<c10::complex<float>> {
     return ret;
   }
   __m512 abs_() const {
-    return _mm512_sqrt_ps(abs_2_());                // abs     abs
+    auto real = _mm512_moveldup_ps(values);    // real real
+    auto imag = _mm512_movehdup_ps(values);    // imag imag
+    return Sleef_hypotf16_u05(real, imag);     // abs  abs
   }
   Vectorized<c10::complex<float>> abs() const {
     const __m512 real_mask = _mm512_castsi512_ps(_mm512_setr_epi32(0xFFFFFFFF, 0x00000000, 0xFFFFFFFF, 0x00000000,
@@ -706,10 +708,7 @@ template <> class Vectorized<c10::complex<float>> {
     auto abs = abs_();
     auto zero = _mm512_setzero_ps();
     auto mask = _mm512_cmp_ps_mask(abs, zero, _CMP_EQ_OQ);
-    auto abs_val = Vectorized(abs);
-
-    auto div = values / abs_val.values;       // x / abs(x)
-
+    auto div = values / abs;
     return _mm512_mask_blend_ps(mask, div, zero);
   }
   __m512 real_() const {
Original file line number	Diff line number	Diff line change
		@@ -1 +1 @@
		f235d4da06905b35d75879a0a9bc3034ab7385ac
		486b32e5b1a2e2964a7db4c7db25b59d879b8796
-Original file line number
+Diff line change
@@ Expand Up / @@ -420,6 +420,7 @@ @@
       - gchanan
       - ezyang
       - dzhulgakov
+      - malfet
       mandatory_checks_name:
       - EasyCLA
       - Lint
@@ Expand Down @@