[AMD] Add 8-GPU MX35X test running DSR1-MXFP4 model for AMD CI (#13602)

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

@@ -543,6 +543,45 @@ jobs:
         run: |
           bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 12
 
+  unit-test-backend-1-gpu-amd-mi35x:
+    needs: [check-changes, stage-a-test-1-amd]
+    if: |
+      always() &&
+      (
+        (inputs.target_stage == 'unit-test-backend-1-gpu-amd-mi35x') ||
+        (
+          !inputs.target_stage &&
+          (!failure() && !cancelled()) &&
+          ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
+        )
+      )
+    strategy:
+      fail-fast: false
+      matrix:
+        runner: [linux-mi35x-gpu-1]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.pr_head_sha || inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ensure_vram_clear.sh rocm
+
+      - name: Start CI container
+        run: bash scripts/ci/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd_ci_install_dependency.sh
+
+      - name: Run test
+        timeout-minutes: 15
+        run: |
+          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-amd-mi35x
+
   unit-test-backend-2-gpu-amd:
     needs: [check-changes, stage-a-test-1-amd]
     if: |
@@ -631,6 +670,35 @@ jobs:
         run: |
           bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 3600
 
+  unit-test-backend-8-gpu-amd-mi35x:
+    needs: [check-changes, unit-test-backend-2-gpu-amd]
+    if: always() && !failure() && !cancelled() &&
+      ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
+    strategy:
+      fail-fast: false
+      matrix:
+        runner: [linux-mi35x-gpu-8]
+    runs-on: ${{matrix.runner}}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ensure_vram_clear.sh rocm
+
+      - name: Start CI container
+        run: bash scripts/ci/amd_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd_ci_install_dependency.sh
+
+      - name: Run test
+        timeout-minutes: 30
+        run: |
+          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd-mi35x --timeout-per-file 1800
+
   performance-test-1-gpu-part-1-amd:
     needs: [check-changes, stage-a-test-1-amd]
     if: |
@@ -897,8 +965,10 @@ jobs:
         stage-b-test-small-1-gpu-amd,
         stage-b-test-large-2-gpu-amd,
         unit-test-backend-1-gpu-amd,
+        unit-test-backend-1-gpu-amd-mi35x,
         unit-test-backend-2-gpu-amd,
         unit-test-backend-8-gpu-amd,
+        unit-test-backend-8-gpu-amd-mi35x,
         performance-test-1-gpu-part-1-amd,
         performance-test-1-gpu-part-2-amd,
         performance-test-2-gpu-amd,

scripts/ci/amd_ci_exec.sh

@@ -60,9 +60,10 @@ if docker exec \
   "${ENV_ARGS[@]}" \
   ci_sglang "$@"; then
   exit 0
+else
+  FIRST_EXIT_CODE=$?
 fi
 
-FIRST_EXIT_CODE=$?
 echo "First attempt failed with exit code $FIRST_EXIT_CODE"
 echo "Retrying with HF_HUB_OFFLINE=1 (offline mode)..."
 

scripts/ci/amd_ci_install_dependency.sh

@@ -99,10 +99,11 @@ case "${GPU_ARCH}" in
   mi35x)
     echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
     docker exec ci_sglang rm -rf python/pyproject.toml && mv python/pyproject_other.toml python/pyproject.toml
-    install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]" --no-deps # TODO: only for mi35x
+    # Follow the same dependency installation flow as mi30x/mi300/mi325.
+    install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
     # For lmms_evals evaluating MMMU
     docker exec -w / ci_sglang git clone --branch v0.4.1 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
-    install_with_retry docker exec -w /lmms-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e . --no-deps
+    install_with_retry docker exec -w /lmms-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
     ;;
   mi30x|mi300|mi325)
     echo "Runner uses ${GPU_ARCH}; will fetch mi30x image."

test/srt/run_suite.py

@@ -157,6 +157,9 @@ suite_amd = {
         TestFile("test_deepseek_v3_basic.py", 275),
         TestFile("test_deepseek_v3_mtp.py", 275),
     ],
+    "per-commit-8-gpu-amd-mi35x": [
+        TestFile("test_deepseek_r1_mxfp4_8gpu.py", 1800),
+    ],
     "nightly-amd": [
         TestFile("nightly/test_gsm8k_eval_amd.py"),
     ],

test/srt/test_deepseek_r1_mxfp4_8gpu.py

@@ -0,0 +1,165 @@
+import os
+import unittest
+from types import SimpleNamespace
+
+import requests
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.send_one import BenchArgs, send_one_prompt
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+DEEPSEEK_R1_MODEL_PATH = "amd/DeepSeek-R1-MXFP4-Preview"
+SERVER_LAUNCH_TIMEOUT = 1000
+
+
+class TestDeepseekR1MXFP4(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_R1_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "8",
+            "--chunked-prefill-size",
+            "131072",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        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)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-r1-mxfp4)\n" f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], 0.94)
+
+    def test_bs_1_speed(self):
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        _, speed = send_one_prompt(args)
+
+        print(f"{speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-r1-mxfp4)\n" f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(speed, 75)
+
+
+class TestDeepseekR1MXFP4MTP(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_R1_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "8",
+            "--chunked-prefill-size",
+            "131072",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+        if "SGLANG_ENABLE_SPEC_V2" in os.environ:
+            del os.environ["SGLANG_ENABLE_SPEC_V2"]
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        requests.get(self.base_url + "/flush_cache")
+
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        server_info = requests.get(self.base_url + "/get_server_info")
+        avg_spec_accept_length = server_info.json()["internal_states"][0][
+            "avg_spec_accept_length"
+        ]
+        print(f"{avg_spec_accept_length=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-r1-mxfp4 mtp)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+                f"{avg_spec_accept_length=:.2f}\n"
+            )
+            self.assertGreater(metrics["accuracy"], 0.94)
+            self.assertGreater(avg_spec_accept_length, 2.04)
+
+    def test_bs_1_speed(self):
+        args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
+        acc_length, speed = send_one_prompt(args)
+
+        print(f"{acc_length=:.2f} {speed=:.2f}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_bs_1_speed (deepseek-r1-mxfp4 mtp)\n"
+                f"{acc_length=:.2f}\n"
+                f"{speed=:.2f} token/s\n"
+            )
+            self.assertGreater(acc_length, 2.04)
+            self.assertGreater(speed, 150)
+
+
+if __name__ == "__main__":
+    unittest.main()