From c8147cddb925ae6ebba9ed7b954fe2a926d26f7e Mon Sep 17 00:00:00 2001 From: ssjia Date: Fri, 13 Mar 2026 13:33:55 -0700 Subject: [PATCH] [ET-VK][testing] Add Skycastle GPU host test workflow via ECOD Create a new Skycastle workflow that mirrors the existing eureka Android device test workflow but runs on an ECOD NEUTRON_NVIDIA_A10M Linux VM with an NVIDIA GPU. This enables testing the Vulkan backend on real NVIDIA hardware in CI. New files: - gpu_host_utils.sky: Utility functions for ECOD GPU host interaction (reserve, upload, ssh, release), analogous to riot_utils.sky but using ecod CLI instead of adb. - executorch_vulkan_gpu_unit_tests.sky: Main workflow (etvk-gpu-host-tests) that exports models, reserves an ECOD GPU VM, runs all tests (gtest binaries, llama runner, classification, greenscreen, scenex, skin seg), and releases the VM. Modified files: - executorch_vulkan_test_utils.sky: Added build_target_for_host() to build test binaries for x86_64 Linux host (no Android cross-compilation). Differential Revision: [D96233257](https://our.internmc.facebook.com/intern/diff/D96233257/) [ghstack-poisoned] --- backends/vulkan/test/custom_ops/q4gsw_linear.cpp | 7 ------- 1 file changed, 7 deletions(-) diff --git a/backends/vulkan/test/custom_ops/q4gsw_linear.cpp b/backends/vulkan/test/custom_ops/q4gsw_linear.cpp index 2af1488541d..1d1d48615e9 100644 --- a/backends/vulkan/test/custom_ops/q4gsw_linear.cpp +++ b/backends/vulkan/test/custom_ops/q4gsw_linear.cpp @@ -499,13 +499,6 @@ void reference_impl(TestCase& test_case) { } int64_t quantized_linear_flop_calculator(const TestCase& test_case) { - int input_idx = 0; - int weight_idx = 1; - if (test_case.operator_name().find("dq8ca") != std::string::npos) { - input_idx = 0; - weight_idx = 3; // Weight comes after input, input_scale, input_zero_point - } - // Get input and weight dimensions const auto& input_sizes = test_case.inputs()[0].get_tensor_sizes(); const auto& output_sizes = test_case.outputs()[0].get_tensor_sizes();