[AMD] Fix TP2 DeepSeek-R1 nhead=64 MLA decode crash and add nightly coverage (#27188)

Co-authored-by: clintg6 <7388379+clintg6@users.noreply.github.com>

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

@@ -38,6 +38,8 @@ on:
           - nightly-4-gpu-rocm720
           # 2-GPU GLM-5.1-MXFP4 (MI35x only)
           - nightly-2-gpu-mi35x-glm51-mxfp4-rocm720
+          # 2-GPU DeepSeek-R1-MXFP4 TP2 (MI35x only)
+          - nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2-rocm720
           # 8-GPU GPT-OSS (MI30x mixes Grok1-FP8; MI35x mixes Qwen3-Coder-Next)
           - nightly-accuracy-8-gpu-rocm720
           - nightly-accuracy-8-gpu-mi35x-rocm720
@@ -61,6 +63,7 @@ on:
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-rocm720
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8-rocm720
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion-rocm720
+          - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720
           - nightly-8-gpu-mi35x-deepseek-r1-hicache-rocm720
           # 8-GPU DeepSeek-V4 (MI35x only)
           - nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720
@@ -262,6 +265,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2-rocm720:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2-rocm720,'))
+    runs-on: linux-mi35x-gpu-2
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - 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-test-time-deps
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x ROCm 7.2 (2-GPU DeepSeek-R1-MXFP4 TP2)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md
+          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-2-gpu-mi35x-deepseek-r1-mxfp4-tp2 --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   nightly-accuracy-2-gpu-vlm-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu-vlm-rocm720,'))
     runs-on: linux-mi325-2gpu-sglang
@@ -1023,6 +1060,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720,'))
+    runs-on: linux-mi35x-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - 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-test-time-deps
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x ROCm 7.2 (DeepSeek-R1-MXFP4 TP4)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md
+          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-deepseek-r1-mxfp4-tp4 --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8-rocm720:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8-rocm720,'))
     runs-on: linux-mi35x-gpu-8
@@ -1717,6 +1788,7 @@ jobs:
       - nightly-perf-2-gpu-vlm-rocm720
       - nightly-4-gpu-rocm720
       - nightly-2-gpu-mi35x-glm51-mxfp4-rocm720
+      - nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2-rocm720
       # 8-GPU GPT-OSS (MI30x mixes Grok1-FP8; MI35x mixes Qwen3-Coder-Next)
       - nightly-accuracy-8-gpu-rocm720
       - nightly-accuracy-8-gpu-mi35x-rocm720
@@ -1740,6 +1812,7 @@ jobs:
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-rocm720
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8-rocm720
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion-rocm720
+      - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720
       - nightly-8-gpu-mi35x-deepseek-r1-hicache-rocm720
       # 8-GPU DeepSeek-V4 (MI35x only)
       - nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720

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

@@ -38,6 +38,8 @@ on:
           - nightly-4-gpu
           # 2-GPU GLM-5.1-MXFP4 (MI35x only)
           - nightly-2-gpu-mi35x-glm51-mxfp4
+          # 2-GPU DeepSeek-R1-MXFP4 TP2 (MI35x only)
+          - nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2
           # 8-GPU GPT-OSS (MI30x mixes Grok1-FP8; MI35x mixes Qwen3-Coder-Next)
           - nightly-accuracy-8-gpu
           - nightly-accuracy-8-gpu-mi35x
@@ -61,6 +63,7 @@ on:
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8
           - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion
+          - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4
           - nightly-8-gpu-mi35x-deepseek-r1-hicache
           # 8-GPU Kimi-K2.6 (MI30x + MI35x)
           - nightly-8-gpu-kimi-k26
@@ -264,6 +267,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2,'))
+    runs-on: linux-mi35x-gpu-2
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - 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
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x (2-GPU DeepSeek-R1-MXFP4 TP2)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md
+          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-2-gpu-mi35x-deepseek-r1-mxfp4-tp2 --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   nightly-accuracy-2-gpu-vlm:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-accuracy-2-gpu-vlm,'))
     runs-on: linux-mi325-2gpu-sglang
@@ -1028,6 +1065,40 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4:
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4,'))
+    runs-on: linux-mi35x-gpu-8
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - 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
+          bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
+
+      - name: Accuracy Test MI35x (DeepSeek-R1-MXFP4 TP4)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md
+          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-deepseek-r1-mxfp4-tp4 --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
   nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8,'))
     runs-on: linux-mi35x-gpu-8
@@ -1648,6 +1719,7 @@ jobs:
       # - nightly-perf-2-gpu-vlm  # excluded: perf failures don't block CI
       - nightly-4-gpu
       - nightly-2-gpu-mi35x-glm51-mxfp4
+      - nightly-2-gpu-mi35x-deepseek-r1-mxfp4-tp2
       # 8-GPU GPT-OSS (MI30x mixes Grok1-FP8; MI35x mixes Qwen3-Coder-Next)
       - nightly-accuracy-8-gpu
       - nightly-accuracy-8-gpu-mi35x
@@ -1671,6 +1743,7 @@ jobs:
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-kv-fp8
       - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion
+      - nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4
       - nightly-8-gpu-mi35x-deepseek-r1-hicache
       # 8-GPU Kimi-K2.6 (MI30x + MI35x)
       - nightly-8-gpu-kimi-k26

python/sglang/srt/layers/attention/aiter_backend.py

@@ -281,7 +281,7 @@ class AiterAttnBackend(AttentionBackend):
             # so all num_head size does not use qh16 kernel to simulate
             # it should not use fake-nps (fast_mode = False, intra_batch_mode = True)
             # it will cause gpu-fault or accuracy issue
-            if self.num_head == 32 or self.num_head == 128:
+            if self.num_head in (32, 64, 128):
                 fast_mode = True
                 intra_batch_mode = False
 

test/registered/amd/accuracy/mi35x/test_deepseek_r1_mxfp4_tp2_mi35x.py

@@ -0,0 +1,185 @@
+"""MI35x DeepSeek-R1-MXFP4 TP=2 GSM8K AITER MLA regression.
+
+DeepSeek-R1 has 128 attention heads, so TP=2 gives 64 heads per rank. This
+covers the AITER persistent MLA decode metadata path that previously crashed
+with GPU memory access faults.
+
+Registry: nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2 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 typing import 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_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+register_amd_ci(
+    est_time=1800,
+    suite="nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2",
+    nightly=True,
+)
+
+INVALID = -9999999
+
+DEEPSEEK_R1_MXFP4_LOCAL_PATH = "/data2/models/amd-DeepSeek-R1-MXFP4-Preview"
+DEEPSEEK_R1_MXFP4_HF_MODEL_ID = "amd/DeepSeek-R1-MXFP4-Preview"
+SERVER_LAUNCH_TIMEOUT = 3600
+GSM8K_ACCURACY_THRESHOLD = 0.93
+
+
+def get_model_path() -> str:
+    env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
+    if env_path:
+        return env_path
+    if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
+        return DEEPSEEK_R1_MXFP4_LOCAL_PATH
+    return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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]:
+    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 TestDeepSeekR1MXFP4TP2MI35x(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = get_model_path()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_AITER_MLA_PERSIST"] = "1"
+
+        cls.process = popen_launch_server(
+            model=cls.model,
+            base_url=cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=[
+                "--attention-backend",
+                "aiter",
+                "--tp",
+                "2",
+                "--chunked-prefill-size",
+                "131072",
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                "0.85",
+                "--trust-remote-code",
+                "--kv-cache-dtype",
+                "fp8_e4m3",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+            ],
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        acc, invalid, latency = run_gsm8k_benchmark(
+            self.base_url, num_questions=self.num_questions
+        )
+        print(f"accuracy={acc:.3f} invalid={invalid:.3f} latency={latency:.1f}s")
+
+        if is_in_ci():
+            write_github_step_summary(
+                "### DeepSeek-R1-MXFP4 TP=2 GSM8K (MI35x)\n\n"
+                "| Model | TP | Examples | Accuracy | Invalid | Threshold | Latency |\n"
+                "| ----- | -- | -------- | -------- | ------- | --------- | ------- |\n"
+                f"| {self.model} | 2 | {self.num_questions} | {acc:.3f} | "
+                f"{invalid:.3f} | {GSM8K_ACCURACY_THRESHOLD:.2f} | {latency:.1f}s |\n"
+            )
+
+        self.assertGreaterEqual(acc, GSM8K_ACCURACY_THRESHOLD)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/amd/accuracy/mi35x/test_deepseek_r1_mxfp4_tp4_mi35x.py

@@ -0,0 +1,184 @@
+"""MI35x DeepSeek-R1-MXFP4 TP=4 GSM8K AITER MLA regression.
+
+DeepSeek-R1 has 128 attention heads, so TP=4 gives 32 heads per rank. This
+covers the AITER persistent MLA decode metadata path for the nhead=32 case.
+
+Registry: nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4 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 typing import 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_URL_FOR_TEST,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+from sglang.utils import download_and_cache_file, read_jsonl
+
+register_amd_ci(
+    est_time=1800,
+    suite="nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4",
+    nightly=True,
+)
+
+INVALID = -9999999
+
+DEEPSEEK_R1_MXFP4_LOCAL_PATH = "/data2/models/amd-DeepSeek-R1-MXFP4-Preview"
+DEEPSEEK_R1_MXFP4_HF_MODEL_ID = "amd/DeepSeek-R1-MXFP4-Preview"
+SERVER_LAUNCH_TIMEOUT = 3600
+GSM8K_ACCURACY_THRESHOLD = 0.93
+
+
+def get_model_path() -> str:
+    env_path = os.environ.get("DEEPSEEK_R1_MXFP4_MODEL_PATH")
+    if env_path:
+        return env_path
+    if os.path.exists(DEEPSEEK_R1_MXFP4_LOCAL_PATH):
+        return DEEPSEEK_R1_MXFP4_LOCAL_PATH
+    return DEEPSEEK_R1_MXFP4_HF_MODEL_ID
+
+
+def get_one_example(lines, i, include_answer):
+    ret = "Question: " + lines[i]["question"] + "\nAnswer:"
+    if include_answer:
+        ret += " " + lines[i]["answer"]
+    return ret
+
+
+def get_few_shot_examples(lines, k):
+    ret = ""
+    for i in range(k):
+        ret += get_one_example(lines, i, True) + "\n\n"
+    return ret
+
+
+def get_answer_value(answer_str):
+    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]:
+    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 TestDeepSeekR1MXFP4TP4MI35x(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = get_model_path()
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.num_questions = int(os.environ.get("GSM8K_NUM_QUESTIONS", "200"))
+
+        env = os.environ.copy()
+        env["SGLANG_USE_AITER"] = "1"
+        env["SGLANG_AITER_MLA_PERSIST"] = "1"
+
+        cls.process = popen_launch_server(
+            model=cls.model,
+            base_url=cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=[
+                "--attention-backend",
+                "aiter",
+                "--tp",
+                "4",
+                "--chunked-prefill-size",
+                "131072",
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                "0.85",
+                "--trust-remote-code",
+                "--kv-cache-dtype",
+                "fp8_e4m3",
+                "--model-loader-extra-config",
+                '{"enable_multithread_load": true}',
+            ],
+            env=env,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        acc, invalid, latency = run_gsm8k_benchmark(
+            self.base_url, num_questions=self.num_questions
+        )
+        print(f"accuracy={acc:.3f} invalid={invalid:.3f} latency={latency:.1f}s")
+
+        if is_in_ci():
+            write_github_step_summary(
+                "### DeepSeek-R1-MXFP4 TP=4 GSM8K (MI35x)\n\n"
+                "| Model | TP | Examples | Accuracy | Invalid | Threshold | Latency |\n"
+                "| ----- | -- | -------- | -------- | ------- | --------- | ------- |\n"
+                f"| {self.model} | 4 | {self.num_questions} | {acc:.3f} | "
+                f"{invalid:.3f} | {GSM8K_ACCURACY_THRESHOLD:.2f} | {latency:.1f}s |\n"
+            )
+
+        self.assertGreaterEqual(acc, GSM8K_ACCURACY_THRESHOLD)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/run_suite.py

@@ -114,6 +114,8 @@ NIGHTLY_SUITES = {
         "nightly-amd-1-gpu",
         "nightly-amd-1-gpu-mi35x",
         "nightly-amd-1-gpu-zimage-turbo",
+        "nightly-amd-2-gpu-mi35x-deepseek-r1-mxfp4-tp2",
+        "nightly-amd-8-gpu-mi35x-deepseek-r1-mxfp4-tp4",
         "nightly-amd-4-gpu",
         "nightly-amd-8-gpu",
         "nightly-amd-vlm",