Use Linux A10 Agents for Integration Tests

.pipelines/integration-tests.yml

@@ -15,7 +15,7 @@
 #   ORT GenAI is built from source on each run, once per (os, arch):
 #     - Windows x64: built with --use_cuda (CUDA 12.8, sm_86 for A10). The
 #                    same wheel exercises the CPU, CUDA, and WebGPU lanes.
-#     - Linux x64:   built with --use_cuda (CUDA 12.8, sm_90 for H100)
+#     - Linux x64:   built with --use_cuda (CUDA 12.8, sm_86 for A10)
 #                    inside the manylinux container. The same wheel
 #                    exercises the CPU and CUDA lanes.
 #     - macOS arm64: built without CUDA. The same wheel exercises the CPU

.pipelines/stages/jobs/integration-build-job.yml

@@ -84,7 +84,7 @@ jobs:
               --use_cuda \
               --cuda_home /usr/local/cuda \
               --cmake_extra_defines MANYLINUX=ON \
-              --cmake_extra_defines 'CMAKE_CUDA_ARCHITECTURES=90-real' \
+              --cmake_extra_defines 'CMAKE_CUDA_ARCHITECTURES=86-real' \
               --cmake_extra_defines PYTHON_EXECUTABLE=/opt/python/cp$(py_no_dot_ver)-cp$(py_no_dot_ver)/bin/python$(py_dot_ver) \
               --skip_tests \
               --skip_examples

.pipelines/stages/jobs/integration-test-job.yml

@@ -25,7 +25,7 @@ jobs:
   displayName: '${{ parameters.ep }} · ${{ parameters.model }}'
   ${{ if eq(parameters.os, 'linux') }}:
     ${{ if eq(parameters.ep, 'cuda') }}:
-      pool: 'onnx-publish-Linux-GPU-H100'
+      pool: 'onnxruntime-Linux-GPU-A10'
     ${{ else }}:
       pool: 'onnxruntime-Ubuntu2204-AMD-CPU'
   ${{ if eq(parameters.os, 'win') }}:

test/python/integration/test_integration_text.py

@@ -29,11 +29,13 @@
 
 # Known (platform, device, model) combinations that don't fit on the
 # agent's GPU memory. TODO: re-enable these once the GPU agents have
-# more VRAM. The current Windows CUDA pool
-# (onnxruntime-Win2022-GPU-A10) only exposes ~4 GB to the job.
+# more VRAM. The Windows CUDA pool (onnxruntime-Win2022-GPU-A10) and the
+# Linux CUDA pool (onnxruntime-Linux-GPU-A10) only expose ~4 GB to the job.
 _VRAM_CONSTRAINED_SKIPS: set[tuple[str, str, str]] = {
     ("win32", "cuda", "ministral-3-3b-Instruct-2512"),
     ("win32", "cuda", "Phi-4-mini-instruct"),
+    ("linux", "cuda", "ministral-3-3b-Instruct-2512"),
+    ("linux", "cuda", "Phi-4-mini-instruct"),
 }