[AMD] [GLM-5 Day 0] Add GLM-5 nightly test (#18911)

.github/workflows/nightly-test-amd-rocm720.yml

@@ -32,6 +32,7 @@ on:
           - 'nightly-8-gpu-deepseek-v32-rocm720'
           - 'nightly-8-gpu-deepseek-v32-mtp-rocm720'
           - 'nightly-8-gpu-kimi-k25-rocm720'
+          - 'nightly-8-gpu-glm5-rocm720'
           # MI35x ROCm 7.2 jobs
           - 'nightly-test-1-gpu-mi35x-rocm720'
           - 'nightly-accuracy-8-gpu-mi35x-rocm720'
@@ -43,6 +44,7 @@ on:
           - 'nightly-perf-8-gpu-mi35x-deepseek-v32-basic-rocm720'
           - 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720'
           - 'nightly-8-gpu-mi35x-kimi-k25-rocm720'
+          - 'nightly-8-gpu-mi35x-glm5-rocm720'
   workflow_call:
     inputs:
       ref:
@@ -494,6 +496,39 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # 8-GPU GLM-5 (Accuracy) ROCm 7.2
+  nightly-8-gpu-glm5-rocm720:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-glm5-rocm720')
+    runs-on: linux-mi325-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker (ROCm 7.2)
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-aiter-build --skip-test-time-deps
+          # GLM-5 requires latest transformers for glm_moe_dsa architecture
+          bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git
+
+      - name: Accuracy Test ROCm 7.2 (8-GPU GLM-5 NSA)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-glm5 --nightly --timeout-per-file 3600 --continue-on-error || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # ============================================== MI35x ROCm 7.2 Tests ==============================================
   # MI35x 1-GPU ROCm 7.2 tests
   nightly-test-1-gpu-mi35x-rocm720:
@@ -825,6 +860,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-8-gpu-mi35x-glm5-rocm720:
+    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-glm5-rocm720')
+    runs-on: linux-mi35x-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Setup docker (ROCm 7.2)
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-aiter-build --skip-test-time-deps
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+          # GLM-5 requires latest transformers for glm_moe_dsa architecture
+          bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git
+
+      - name: Accuracy Test MI35x ROCm 7.2 (8-GPU GLM-5 NSA)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-glm5 --nightly --timeout-per-file 7200 --continue-on-error || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP) ROCm 7.2
   nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720')
@@ -877,6 +946,7 @@ jobs:
       - nightly-8-gpu-deepseek-v32-rocm720
       - nightly-8-gpu-deepseek-v32-mtp-rocm720
       - nightly-8-gpu-kimi-k25-rocm720
+      - nightly-8-gpu-glm5-rocm720
       # MI35x ROCm 7.2 jobs
       - nightly-test-1-gpu-mi35x-rocm720
       - nightly-accuracy-8-gpu-mi35x-rocm720
@@ -888,6 +958,7 @@ jobs:
       - nightly-perf-8-gpu-mi35x-deepseek-v32-basic-rocm720
       - nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720
       - nightly-8-gpu-mi35x-kimi-k25-rocm720
+      - nightly-8-gpu-mi35x-glm5-rocm720
     runs-on: ubuntu-latest
     steps:
       - name: Check if any job failed

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

@@ -32,9 +32,11 @@ on:
           - 'nightly-8-gpu-deepseek-v32'
           - 'nightly-8-gpu-deepseek-v32-mtp'
           - 'nightly-8-gpu-kimi-k25'
+          - 'nightly-8-gpu-glm5'
           # MI35x jobs
           - 'nightly-test-1-gpu-mi35x'
           - 'nightly-8-gpu-mi35x-kimi-k25'
+          - 'nightly-8-gpu-mi35x-glm5'
           - 'nightly-accuracy-8-gpu-mi35x'
           - 'nightly-8-gpu-mi35x-grok1-int4'
           - 'nightly-8-gpu-mi35x-grok2'
@@ -494,6 +496,38 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-8-gpu-glm5:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-8-gpu-glm5')
+    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: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd/amd_ci_install_dependency.sh
+          # GLM-5 requires latest transformers for glm_moe_dsa architecture
+          bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git
+
+      - name: Accuracy Test (8-GPU GLM-5 NSA)
+        timeout-minutes: 120
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-glm5 --nightly --timeout-per-file 3600 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # ============================================== MI35x Tests ==============================================
   # MI35x 1-GPU tests - platform-agnostic tests that may work on CDNA4 (gfx950)
   nightly-test-1-gpu-mi35x:
@@ -827,6 +861,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-8-gpu-mi35x-glm5:
+    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-glm5')
+    runs-on: linux-mi35x-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: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: |
+          bash scripts/ci/amd/amd_ci_install_dependency.sh
+          # Install tabulate for run_suite.py (missing in MI35x container)
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+          # GLM-5 requires latest transformers for glm_moe_dsa architecture
+          bash scripts/ci/amd/amd_ci_exec.sh pip install git+https://github.com/huggingface/transformers.git
+
+      - name: Accuracy Test MI35x (8-GPU GLM-5 NSA)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-glm5 --nightly --timeout-per-file 7200 || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # MI35x 8-GPU DeepSeek-V3.2 Performance Test (MTP)
   nightly-perf-8-gpu-mi35x-deepseek-v32-mtp:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (inputs.job_filter == '' || inputs.job_filter == 'all' || inputs.job_filter == 'nightly-perf-8-gpu-mi35x-deepseek-v32-mtp')
@@ -879,6 +947,7 @@ jobs:
       - nightly-8-gpu-deepseek-v32
       - nightly-8-gpu-deepseek-v32-mtp
       - nightly-8-gpu-kimi-k25
+      - nightly-8-gpu-glm5
       # MI35x jobs
       - nightly-test-1-gpu-mi35x
       - nightly-accuracy-8-gpu-mi35x
@@ -888,6 +957,7 @@ jobs:
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp
       - nightly-8-gpu-mi35x-kimi-k25
+      - nightly-8-gpu-mi35x-glm5
       # 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

test/registered/amd/accuracy/mi30x/test_glm5_eval_amd.py

@@ -0,0 +1,244 @@
+"""AMD GLM-5 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests GLM-5 with NSA attention backend using few-shot completion
+benchmark on MI325/MI300X.
+
+Registry: nightly-amd-accuracy-8-gpu-glm5 suite
+"""
+
+import ast
+import os
+import re
+import time
+import unittest
+from dataclasses import dataclass, field
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+# Register for AMD CI - GLM-5 accuracy test (~60 min)
+register_amd_ci(
+    est_time=3600,
+    suite="nightly-amd-accuracy-8-gpu-glm5",
+    nightly=True,
+)
+
+INVALID = -9999999
+
+
+@dataclass
+class ModelConfig:
+    """Configuration for a model to test."""
+
+    model_path: str
+    tp_size: int = 8
+    accuracy_threshold: float = 0.50
+    other_args: List[str] = field(default_factory=list)
+    env_vars: dict = field(default_factory=dict)
+    timeout: Optional[int] = None
+    variant: Optional[str] = None
+
+    def get_display_name(self) -> str:
+        if self.variant:
+            return f"{self.model_path} ({self.variant})"
+        return self.model_path
+
+
+# GLM-5 models for MI325/MI300X - NSA attention backend
+GLM5_MODELS = [
+    # GLM-5 with NSA attention (TP=8)
+    ModelConfig(
+        model_path="zai-org/GLM-5",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=3600,
+        variant="nsa",
+        other_args=[
+            "--trust-remote-code",
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.80",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",  # 20 minutes for weight loading
+        ],
+        env_vars={"SGLANG_USE_AITER": "1"},
+    ),
+]
+
+
+def get_one_example(lines, i, include_answer):
+    """Format a single GSM8K example."""
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    """Get k few-shot examples for prompting."""
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    """Extract numerical answer from response."""
+    answer_str = answer_str.replace(",", "")
+    numbers = re.findall(r"\d+", answer_str)
+    if len(numbers) < 1:
+        return INVALID
+    try:
+        return ast.literal_eval(numbers[-1])
+    except SyntaxError:
+        return INVALID
+
+
+def run_gsm8k_benchmark(
+    base_url: str,
+    num_questions: int = 200,
+    num_shots: int = 5,
+    parallel: int = 64,
+) -> Tuple[float, float, float]:
+    """Run GSM8K few-shot completion benchmark."""
+    import sglang as sgl
+    from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
+
+    url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
+    data_path = download_and_cache_file(url)
+    lines = list(read_jsonl(data_path))
+
+    few_shot_examples = get_few_shot_examples(lines, num_shots)
+
+    questions = []
+    labels = []
+    for i in range(len(lines[:num_questions])):
+        questions.append(get_one_example(lines, i, False))
+        labels.append(get_answer_value(lines[i]["answer"]))
+    assert all(l != INVALID for l in labels)
+    arguments = [{"question": q} for q in questions]
+
+    @sgl.function
+    def few_shot_gsm8k(s, question):
+        s += few_shot_examples + question
+        s += sgl.gen(
+            "answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
+        )
+
+    backend = RuntimeEndpoint(base_url)
+    sgl.set_default_backend(backend)
+
+    tic = time.perf_counter()
+    states = few_shot_gsm8k.run_batch(
+        arguments, temperature=0, num_threads=parallel, progress_bar=True
+    )
+    latency = time.perf_counter() - tic
+
+    preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
+    acc = np.mean(np.array(preds) == np.array(labels))
+    invalid = np.mean(np.array(preds) == INVALID)
+
+    return float(acc), float(invalid), float(latency)
+
+
+class TestGLM5EvalAMD(unittest.TestCase):
+    """GLM-5 GSM8K Completion Evaluation Test for AMD MI325/MI300X."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = GLM5_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_glm5_accuracy(self):
+        """Test GLM-5 models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### GLM-5 Models (MI325)\n\n"
+        summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            display_name = config.get_display_name()
+            with self.subTest(model=display_name):
+                print(f"\n{'='*60}")
+                print(f"Testing: {display_name}")
+                print(f"{'='*60}")
+
+                env = os.environ.copy()
+                for key, value in config.env_vars.items():
+                    env[key] = value
+
+                other_args = list(config.other_args)
+                other_args.extend(["--tp", str(config.tp_size)])
+                timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+                try:
+                    process = popen_launch_server(
+                        model=config.model_path,
+                        base_url=self.base_url,
+                        timeout=timeout,
+                        other_args=other_args,
+                        env=env,
+                    )
+
+                    try:
+                        acc, invalid, latency = run_gsm8k_benchmark(
+                            self.base_url, num_questions=self.num_questions
+                        )
+                        passed = acc >= config.accuracy_threshold
+                        status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": display_name,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": display_name,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi30x/test_gsm8k_eval_amd.py

@@ -42,7 +42,7 @@ MODEL_SCORE_THRESHOLDS = {
     "meta-llama/Llama-3.2-3B-Instruct": 0.55,
     # Mistral series
     "mistralai/Mistral-7B-Instruct-v0.3": 0.55,
-    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.61,
+    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.58,
     # DeepSeek series
     "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.85,
     # Qwen2 series

test/registered/amd/accuracy/mi35x/test_glm5_eval_mi35x.py

@@ -0,0 +1,249 @@
+"""MI35x GLM-5 GSM8K Completion Evaluation Test (8-GPU)
+
+Tests GLM-5 with NSA attention backend using few-shot completion
+benchmark on MI35x.
+
+Registry: nightly-amd-8-gpu-mi35x-glm5 suite
+"""
+
+import ast
+import os
+
+# Set HF cache for MI35x
+os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
+os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
+
+import re
+import time
+import unittest
+from dataclasses import dataclass, field
+from typing import List, Optional, Tuple
+
+import numpy as np
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+# Register for AMD CI - MI35x GLM-5 accuracy test (~90 min)
+register_amd_ci(
+    est_time=5400,
+    suite="nightly-amd-8-gpu-mi35x-glm5",
+    nightly=True,
+)
+
+INVALID = -9999999
+
+
+@dataclass
+class ModelConfig:
+    """Configuration for a model to test."""
+
+    model_path: str
+    tp_size: int = 8
+    accuracy_threshold: float = 0.50
+    other_args: List[str] = field(default_factory=list)
+    env_vars: dict = field(default_factory=dict)
+    timeout: Optional[int] = None
+    variant: Optional[str] = None
+
+    def get_display_name(self) -> str:
+        if self.variant:
+            return f"{self.model_path} ({self.variant})"
+        return self.model_path
+
+
+# GLM-5 models for MI35x - NSA attention backend
+MI35X_GLM5_MODELS = [
+    # GLM-5 with NSA attention (TP=8)
+    ModelConfig(
+        model_path="zai-org/GLM-5",
+        tp_size=8,
+        accuracy_threshold=0.93,
+        timeout=5400,
+        variant="nsa",
+        other_args=[
+            "--trust-remote-code",
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
+            "--chunked-prefill-size",
+            "131072",
+            "--mem-fraction-static",
+            "0.80",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--watchdog-timeout",
+            "1200",  # 20 minutes for weight loading
+        ],
+        env_vars={},
+    ),
+]
+
+
+def get_one_example(lines, i, include_answer):
+    """Format a single GSM8K example."""
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    """Get k few-shot examples for prompting."""
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    """Extract numerical answer from response."""
+    answer_str = answer_str.replace(",", "")
+    numbers = re.findall(r"\d+", answer_str)
+    if len(numbers) < 1:
+        return INVALID
+    try:
+        return ast.literal_eval(numbers[-1])
+    except SyntaxError:
+        return INVALID
+
+
+def run_gsm8k_benchmark(
+    base_url: str,
+    num_questions: int = 200,
+    num_shots: int = 5,
+    parallel: int = 64,
+) -> Tuple[float, float, float]:
+    """Run GSM8K few-shot completion benchmark."""
+    import sglang as sgl
+    from sglang.lang.backend.runtime_endpoint import RuntimeEndpoint
+
+    url = "https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl"
+    data_path = download_and_cache_file(url)
+    lines = list(read_jsonl(data_path))
+
+    few_shot_examples = get_few_shot_examples(lines, num_shots)
+
+    questions = []
+    labels = []
+    for i in range(len(lines[:num_questions])):
+        questions.append(get_one_example(lines, i, False))
+        labels.append(get_answer_value(lines[i]["answer"]))
+    assert all(l != INVALID for l in labels)
+    arguments = [{"question": q} for q in questions]
+
+    @sgl.function
+    def few_shot_gsm8k(s, question):
+        s += few_shot_examples + question
+        s += sgl.gen(
+            "answer", max_tokens=512, stop=["Question", "Assistant:", "<|separator|>"]
+        )
+
+    backend = RuntimeEndpoint(base_url)
+    sgl.set_default_backend(backend)
+
+    tic = time.perf_counter()
+    states = few_shot_gsm8k.run_batch(
+        arguments, temperature=0, num_threads=parallel, progress_bar=True
+    )
+    latency = time.perf_counter() - tic
+
+    preds = [get_answer_value(states[i]["answer"]) for i in range(len(states))]
+    acc = np.mean(np.array(preds) == np.array(labels))
+    invalid = np.mean(np.array(preds) == INVALID)
+
+    return float(acc), float(invalid), float(latency)
+
+
+class TestGLM5EvalMI35x(unittest.TestCase):
+    """GLM-5 GSM8K Completion Evaluation Test for AMD MI35x."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.models = MI35X_GLM5_MODELS
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+    def test_glm5_accuracy(self):
+        """Test GLM-5 models with GSM8K completion benchmark."""
+        all_results = []
+        summary = "### GLM-5 Models (MI35x)\n\n"
+        summary += "| Model | Variant | TP | Accuracy | Threshold | Status |\n"
+        summary += "| ----- | ------- | -- | -------- | --------- | ------ |\n"
+
+        for config in self.models:
+            display_name = config.get_display_name()
+            with self.subTest(model=display_name):
+                print(f"\n{'='*60}")
+                print(f"Testing: {display_name}")
+                print(f"{'='*60}")
+
+                env = os.environ.copy()
+                for key, value in config.env_vars.items():
+                    env[key] = value
+
+                other_args = list(config.other_args)
+                other_args.extend(["--tp", str(config.tp_size)])
+                timeout = config.timeout or DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+
+                try:
+                    process = popen_launch_server(
+                        model=config.model_path,
+                        base_url=self.base_url,
+                        timeout=timeout,
+                        other_args=other_args,
+                        env=env,
+                    )
+
+                    try:
+                        acc, invalid, latency = run_gsm8k_benchmark(
+                            self.base_url, num_questions=self.num_questions
+                        )
+                        passed = acc >= config.accuracy_threshold
+                        status = "✅ PASS" if passed else "❌ FAIL"
+                        print(
+                            f"  accuracy={acc:.3f} threshold={config.accuracy_threshold} {status}"
+                        )
+
+                        all_results.append(
+                            {
+                                "model": display_name,
+                                "accuracy": acc,
+                                "passed": passed,
+                            }
+                        )
+                        summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | {acc:.3f} | {config.accuracy_threshold} | {status} |\n"
+
+                    finally:
+                        kill_process_tree(process.pid)
+
+                except Exception as e:
+                    summary += f"| {config.model_path} | {config.variant or 'N/A'} | {config.tp_size} | N/A | {config.accuracy_threshold} | ❌ ERROR |\n"
+                    all_results.append(
+                        {
+                            "model": display_name,
+                            "accuracy": None,
+                            "passed": False,
+                            "error": str(e),
+                        }
+                    )
+
+        if is_in_ci():
+            write_github_step_summary(summary)
+
+        failed = [r for r in all_results if not r["passed"]]
+        if failed:
+            raise AssertionError(f"Failed models: {[r['model'] for r in failed]}")
+
+
+if __name__ == "__main__":
+    unittest.main()