[Intel XPU] Initially add nightly GSM8K accuracy tests for Llama-3.1-8B (TP=2) and Qwen3-32B (TP=4) (#28908)

Co-authored-by: Singhal, Shubham <shubham.singhal@intel.com>

test/registered/xpu/llm_models/test_xpu_llama_3_1_8b.py

@@ -0,0 +1,39 @@
+"""Llama-3.1-8B-Instruct GSM8K accuracy on Intel XPU (TP=2).
+
+TP=4 wedges the Level Zero driver during the first prefill batch on Arc/BMG;
+TP=2 runs cleanly with the same model and serves at ~18 tok/s.
+
+Scored by ``simple_eval_gsm8k.GSM8KEval`` (the same evaluator AMD and
+NVIDIA nightlies use); threshold mirrors theirs.
+"""
+
+import unittest
+
+import torch
+
+from sglang.test.ci.ci_register import register_xpu_ci
+from sglang.test.test_utils import CustomTestCase
+from sglang.test.xpu.simple_eval_gsm8k_xpu_mixin import SimpleEvalGSM8KXPUMixin
+
+register_xpu_ci(est_time=1200, suite="nightly-xpu-2-gpu", nightly=True)
+
+
+@unittest.skipUnless(
+    torch.xpu.is_available(),
+    "Intel XPU not available (torch.xpu.is_available() returned False)",
+)
+class TestLlama31_8BInstructXPU(SimpleEvalGSM8KXPUMixin, CustomTestCase):
+    model = "meta-llama/Llama-3.1-8B-Instruct"
+    tp_size = 2
+    accuracy = 0.80
+
+    other_args = SimpleEvalGSM8KXPUMixin.other_args + [
+        "--max-total-tokens",
+        "65536",
+        "--mem-fraction-static",
+        "0.8",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/xpu/llm_models/test_xpu_qwen3_32b.py

@@ -0,0 +1,40 @@
+"""Qwen3-32B GSM8K accuracy on Intel XPU (TP=4).
+
+Scored by ``simple_eval_gsm8k.GSM8KEval`` (the same evaluator AMD and
+NVIDIA nightlies use).
+"""
+
+import unittest
+
+import torch
+
+from sglang.test.ci.ci_register import register_xpu_ci
+from sglang.test.test_utils import CustomTestCase
+from sglang.test.xpu.simple_eval_gsm8k_xpu_mixin import SimpleEvalGSM8KXPUMixin
+
+register_xpu_ci(est_time=1800, suite="nightly-xpu-4-gpu", nightly=True)
+
+
+@unittest.skipUnless(
+    torch.xpu.is_available(),
+    "Intel XPU not available (torch.xpu.is_available() returned False)",
+)
+class TestQwen3_32BXPU(SimpleEvalGSM8KXPUMixin, CustomTestCase):
+    model = "Qwen/Qwen3-32B"
+    tp_size = 4
+    accuracy = 0.85
+    # 64GB BF16 weights split across 4 ranks take ~9 min to load on Intel
+    # Arc Pro B60; the default 600s timeout fires mid-startup. Mirror the
+    # XPU 70B test's 1-hour budget.
+    timeout_for_server_launch = 3600
+
+    other_args = SimpleEvalGSM8KXPUMixin.other_args + [
+        "--max-total-tokens",
+        "65536",
+        "--mem-fraction-static",
+        "0.8",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/run_suite.py

@@ -169,7 +169,11 @@ NIGHTLY_SUITES = {
         "full-8-npu-a3",
         "full-16-npu-a3",
     ],
-    HWBackend.XPU: [],
+    HWBackend.XPU: [
+        "nightly-xpu-1-gpu",
+        "nightly-xpu-2-gpu",
+        "nightly-xpu-4-gpu",
+    ],
 }