[AMD] Fix Janus-Pro crash and add Kimi-K2.5 nightly test (#18269)

.github/workflows/nightly-test-amd.yml

@@ -31,10 +31,10 @@ on:
           - 'nightly-8-gpu-deepseek-v31'
           - 'nightly-8-gpu-deepseek-v32'
           - 'nightly-8-gpu-deepseek-v32-mtp'
-          - 'nightly-8-gpu-kimi-k2'
+          - 'nightly-8-gpu-kimi-k25'
           # MI35x jobs
           - 'nightly-test-1-gpu-mi35x'
-          - 'nightly-8-gpu-mi35x-kimi-k2'
+          - 'nightly-8-gpu-mi35x-kimi-k25'
           - 'nightly-accuracy-8-gpu-mi35x'
           - 'nightly-8-gpu-mi35x-grok1-int4'
           - 'nightly-8-gpu-mi35x-grok2'
@@ -494,6 +494,36 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # 8-GPU Kimi-K2.5 (Accuracy)
+  nightly-8-gpu-kimi-k25:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-kimi-k25')
+    runs-on: linux-mi325-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Accuracy Test (8-GPU Kimi-K2.5)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-kimi-k25 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # ============================================== MI35x Tests ==============================================
   # MI35x 1-GPU tests - platform-agnostic tests that may work on CDNA4 (gfx950)
   nightly-test-1-gpu-mi35x:
@@ -794,9 +824,9 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
-  # MI35x 8-GPU Kimi-K2 (Accuracy)
+  # MI35x 8-GPU Kimi-K2.5 (Accuracy)
-  nightly-8-gpu-mi35x-kimi-k2:
+  nightly-8-gpu-mi35x-kimi-k25:
-    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-mi35x-kimi-k2')
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-mi35x-kimi-k25')
     runs-on: linux-mi35x-gpu-8
     steps:
       - name: Checkout code
@@ -817,13 +847,13 @@ jobs:
           # Install tabulate for run_suite.py (missing in MI35x container)
           bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
 
-      - name: Accuracy Test MI35x (8-GPU Kimi-K2)
+      - name: Accuracy Test MI35x (8-GPU Kimi-K2.5)
         timeout-minutes: 180
         run: |
           > github_summary.md  # Clear summary file
           bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
             -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-mi35x-kimi-k2 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-mi35x-kimi-k25 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
@@ -878,7 +908,7 @@ jobs:
       - nightly-8-gpu-deepseek-v31
       - nightly-8-gpu-deepseek-v32
       - nightly-8-gpu-deepseek-v32-mtp
-      - nightly-8-gpu-kimi-k2
+      - nightly-8-gpu-kimi-k25
       # MI35x jobs
       - nightly-test-1-gpu-mi35x
       - nightly-accuracy-8-gpu-mi35x
@@ -887,7 +917,7 @@ jobs:
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp
-      - nightly-8-gpu-mi35x-kimi-k2
+      - nightly-8-gpu-mi35x-kimi-k25
       # MI35x perf jobs excluded from check - perf failures don't block CI
       # - nightly-perf-8-gpu-mi35x-deepseek-v32-basic
       # - nightly-perf-8-gpu-mi35x-deepseek-v32-mtp

python/sglang/srt/models/deepseek_janus_pro.py

@@ -1955,7 +1955,7 @@ class MultiModalityCausalLM(MultiModalityPreTrainedModel):
         self.language_model = LlamaForCausalLM(
             language_config, quant_config=quant_config
         )
-        self.logits_processor = LogitsProcessor(config)
+        self.logits_processor = LogitsProcessor(language_config)
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
         pixel_values = torch.concat([item.feature for item in items], dim=0)

test/registered/amd/accuracy/mi30x/test_kimi_k25_eval_amd.py

@@ -0,0 +1,104 @@
+"""AMD Kimi-K2.5 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests moonshotai/Kimi-K2.5 with GSM8K few-shot benchmark on MI325.
+
+Registry: nightly-amd-accuracy-8-gpu-kimi-k25 suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - Kimi K2.5 accuracy test (~60 min)
+register_amd_ci(
+    est_time=3600, suite="nightly-amd-accuracy-8-gpu-kimi-k25", nightly=True
+)
+
+KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
+SERVER_LAUNCH_TIMEOUT = 3600
+ACCURACY_THRESHOLD = 0.92
+TP_SIZE = 8
+
+
+class TestKimiK25EvalAMD(CustomTestCase):
+    """Kimi-K2.5 GSM8K Completion Evaluation Test for AMD MI325."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = KIMI_K25_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            str(TP_SIZE),
+            "--decode-attention-backend",
+            "triton",
+            "--prefill-attention-backend",
+            "aiter",
+            "--trust-remote-code",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_ROCM_FUSED_DECODE_MLA"] = "0"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_kimi_k25_gsm8k_accuracy(self):
+        """Test Kimi-K2.5 with GSM8K few-shot completion benchmark."""
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=8,
+            data_path=None,
+            num_questions=1319,
+            parallel=1319,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        acc = metrics["accuracy"]
+
+        passed = acc >= ACCURACY_THRESHOLD
+        status = "✅ PASS" if passed else "❌ FAIL"
+        print(f"  accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")
+
+        if is_in_ci():
+            summary = "### Kimi-K2.5 Model (MI325)\n\n"
+            summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+            summary += "| ----- | -- | -------- | --------- | ------ |\n"
+            summary += f"| {KIMI_K25_MODEL_PATH} | {TP_SIZE} | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
+            write_github_step_summary(summary)
+
+        self.assertGreaterEqual(
+            acc,
+            ACCURACY_THRESHOLD,
+            f"Kimi-K2.5 accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi35x/test_kimi_k25_eval_mi35x.py

@@ -0,0 +1,106 @@
+"""MI35x Kimi-K2.5 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests moonshotai/Kimi-K2.5 with GSM8K few-shot benchmark on MI35x.
+
+Registry: nightly-amd-accuracy-8-gpu-mi35x-kimi-k25 suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+# Register for AMD CI - Kimi K2.5 accuracy test on MI35x (~60 min)
+register_amd_ci(
+    est_time=3600, suite="nightly-amd-accuracy-8-gpu-mi35x-kimi-k25", nightly=True
+)
+
+KIMI_K25_MODEL_PATH = "moonshotai/Kimi-K2.5"
+SERVER_LAUNCH_TIMEOUT = 3600
+ACCURACY_THRESHOLD = 0.92
+TP_SIZE = 8
+
+
+class TestKimiK25EvalMI35x(CustomTestCase):
+    """Kimi-K2.5 GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+    def test_kimi_k25_gsm8k_accuracy(self):
+        """Test Kimi-K2.5 with GSM8K few-shot completion benchmark."""
+        other_args = [
+            "--tp",
+            str(TP_SIZE),
+            "--decode-attention-backend",
+            "triton",
+            "--prefill-attention-backend",
+            "aiter",
+            "--trust-remote-code",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",
+        ]
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_ROCM_FUSED_DECODE_MLA"] = "0"
+
+        process = popen_launch_server(
+            KIMI_K25_MODEL_PATH,
+            self.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+            env=env,
+        )
+
+        try:
+            requests.get(self.base_url + "/flush_cache")
+
+            args = SimpleNamespace(
+                num_shots=8,
+                data_path=None,
+                num_questions=1319,
+                parallel=1319,
+                max_new_tokens=512,
+                host="http://127.0.0.1",
+                port=int(self.base_url.split(":")[-1]),
+            )
+            metrics = run_eval_few_shot_gsm8k(args)
+            acc = metrics["accuracy"]
+
+            passed = acc >= ACCURACY_THRESHOLD
+            status = "✅ PASS" if passed else "❌ FAIL"
+            print(f"  accuracy={acc:.3f} threshold={ACCURACY_THRESHOLD} {status}")
+
+            if is_in_ci():
+                summary = "### Kimi-K2.5 Model (MI35x)\n\n"
+                summary += "| Model | TP | Accuracy | Threshold | Status |\n"
+                summary += "| ----- | -- | -------- | --------- | ------ |\n"
+                summary += f"| {KIMI_K25_MODEL_PATH} | {TP_SIZE} | {acc:.3f} | {ACCURACY_THRESHOLD} | {status} |\n"
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                acc,
+                ACCURACY_THRESHOLD,
+                f"Kimi-K2.5 accuracy {acc:.3f} below threshold {ACCURACY_THRESHOLD}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()