Tags: sadsff/pytorch
Tags
Update on "[dtensor] add backward support for scaled dot product atte… …ntion (flash-attention)" As titled, as a followup to the forward part pytorch#120298. cc mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu wanchaol fduwjj wz337 wconstab yf225 chauhang [ghstack-poisoned]
CUDAEvent::elapsed_time could accidentally initialize a non-used GPU This sets the device before call cudaEventElapsedTime to avoid the case where the "cudaGetCurrentDevice" device would be initialized even though neither event is on that device. [ghstack-poisoned]
[export][reland] Fix unflattened submodule ordering. (pytorch#122507) Summary: Make sure the order of submodules is the same as the original eager module. bypass-github-export-checks Test Plan: buck test mode/opt caffe2/test:test_export -- -r test_unflatten_submodule_ordering buck run mode/opt aps_models/ads/icvr/tests:export_test -- -r test_500x_icvr_fm_export Reviewed By: tugsbayasgalan Differential Revision: D55251277
[Inductor] Fix unbacked symbol in stride when using item() Co-authored-by: Edward Z. Yang <ezyang@meta.com>
Rewrite quantized conv transpose2d for vulkan Summary: Vulkan rewrite sp that quantized transpose 2d ops can run in a model Test Plan: Run vulkan api test: # buck2 build --target-platforms ovr_config//platform/macos:arm64-fbsource //xplat/caffe2:pt_vulkan_api_test_binAppleMac\#macosx-arm64 -c pt.vulkan_full_precision=1 --show-output" # buck-out//v2/gen/fbsource/xplat/caffe2/pt_vulkan_api_test_binAppleMac Running main() from third-party/googletest/1.14.0/googletest/googletest/src/gtest_main.cc [==========] Running 418 tests from 1 test suite. [----------] Global test environment set-up. [----------] 418 tests from VulkanAPITest .... [----------] Global test environment tear-down [==========] 418 tests from 1 test suite ran. (4510 ms total) [ PASSED ] 417 tests. [ SKIPPED ] 1 test, listed below: [ SKIPPED ] VulkanAPITest.querypool_flushed_shader_log YOU HAVE 9 DISABLED TESTS Run quantized vulkan api test: Note the linear quantized are failing but all the convolution tests still pass. Linear failures are being debugged. # buck2 build --target-platforms ovr_config//platform/macos:arm64-fbsource //xplat/caffe2:pt_vulkan_quantized_api_test_binAppleMac\#macosx-arm64 -c pt.vulkan_full_precision=1 --show-output" # buck-out//v2/gen/fbsource/xplat/caffe2/pt_vulkan_quantized_api_test_binAppleMac Running main() from third-party/googletest/1.14.0/googletest/googletest/src/gtest_main.cc [==========] Running 86 tests from 1 test suite. [----------] Global test environment set-up. [----------] 86 tests from VulkanAPITest ... [ PASSED ] 77 tests. [ FAILED ] 9 tests, listed below: [ FAILED ] VulkanAPITest.linear_2d_flat [ FAILED ] VulkanAPITest.linear_2d_small [ FAILED ] VulkanAPITest.linear_2d_large [ FAILED ] VulkanAPITest.linear_3d_flat [ FAILED ] VulkanAPITest.linear_3d_small [ FAILED ] VulkanAPITest.linear_3d_large [ FAILED ] VulkanAPITest.linear_4d_flat [ FAILED ] VulkanAPITest.linear_4d_small [ FAILED ] VulkanAPITest.linear_4d_large 9 FAILED TESTS YOU HAVE 8 DISABLED TESTS # Run CUNET quantized model on hibiki board. Reviewed By: manuelcandales Differential Revision: D52344263
PreviousNext