[AMD] Add Qwen3.8-Flash-Next-FP8 nightly validation on ROCm 7.2 (verified on rocm720)

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

@@ -41,6 +41,8 @@ on:
           - nightly-test-1-gpu-unit-rocm720
           - nightly-test-1-gpu-kernel-rocm720
           - nightly-test-1-gpu-mi35x-rocm720
+          - nightly-2-gpu-mi30x-qwen38-flash-next-fp8-rocm720
+          - nightly-1-gpu-mi35x-qwen38-flash-next-fp8-rocm720
           # 2-GPU and 4-GPU Tests (MI30x + MI35x)
           - nightly-accuracy-2-gpu-rocm720
           - nightly-accuracy-2-gpu-vlm-rocm720
@@ -266,6 +268,101 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
+  # ==============================================================================
+  # Qwen3.8-Flash-Next-FP8 (MI30x TP2+EP2, MI35x TP1)
+  # ==============================================================================
+
+  nightly-2-gpu-mi30x-qwen38-flash-next-fp8-rocm720:
+    name: ${{ format('nightly-2-gpu-mi30x-qwen38-flash-next-fp8 ({0}, linux-mi300-2gpu-sglang)', matrix.rocm_version) }}
+    strategy:
+      fail-fast: false
+      matrix:
+        rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm724", "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-mi30x-qwen38-flash-next-fp8-rocm720,'))
+    runs-on: linux-mi300-2gpu-sglang
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - name: Setup docker (${{ matrix.rocm_version }})
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+          ENABLE_CACHE_HOST: "1"
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Check model cache space
+        run: |
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout \
+            bash scripts/ci/amd/check_hf_cache_space.sh \
+            Qwen/Qwen3.8-Flash-Next-FP8 250
+
+      - name: Qwen3.8-Flash-Next-FP8 Accuracy MI30x ROCm 7.2
+        timeout-minutes: 60
+        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" \
+            -e QWEN38_FLASH_NEXT_FP8_TP_SIZE=2 \
+            -e QWEN38_FLASH_NEXT_FP8_EP_SIZE=2 \
+            -e QWEN38_FLASH_NEXT_FP8_MEM_FRACTION_STATIC=0.75 \
+            python3 run_suite.py --hw amd --suite nightly-amd-2-gpu-mi30x-qwen38-flash-next-fp8 --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md)" >> "$GITHUB_STEP_SUMMARY" || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  nightly-1-gpu-mi35x-qwen38-flash-next-fp8-rocm720:
+    name: ${{ format('nightly-1-gpu-mi35x-qwen38-flash-next-fp8 ({0}, linux-mi35x-gpu-1)', matrix.rocm_version) }}
+    strategy:
+      fail-fast: false
+      matrix:
+        rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm724", "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-1-gpu-mi35x-qwen38-flash-next-fp8-rocm720,'))
+    runs-on: linux-mi35x-gpu-1
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref || github.sha }}
+
+      - name: Ensure VRAM is clear
+        run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
+
+      - name: Setup docker (${{ matrix.rocm_version }})
+        run: |
+          touch github_summary.md
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+          ENABLE_CACHE_HOST: "1"
+
+      - name: Install dependencies
+        run: bash scripts/ci/amd/amd_ci_install_dependency.sh
+
+      - name: Check model cache space
+        run: |
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout \
+            bash scripts/ci/amd/check_hf_cache_space.sh \
+            Qwen/Qwen3.8-Flash-Next-FP8 250
+
+      - name: Qwen3.8-Flash-Next-FP8 Accuracy MI35x ROCm 7.2
+        timeout-minutes: 60
+        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-1-gpu-mi35x-qwen38-flash-next-fp8 --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md)" >> "$GITHUB_STEP_SUMMARY" || true
+          exit ${TEST_EXIT_CODE:-0}
+
   # ==============================================================================
   # 2-GPU and 4-GPU Tests (MI30x + MI35x)
   # ==============================================================================
@@ -2317,6 +2414,8 @@ jobs:
       # 1-GPU Unit Tests (MI30x + MI35x)
       - nightly-test-1-gpu-unit-rocm720
       - nightly-test-1-gpu-mi35x-rocm720
+      - nightly-2-gpu-mi30x-qwen38-flash-next-fp8-rocm720
+      - nightly-1-gpu-mi35x-qwen38-flash-next-fp8-rocm720
       # 2-GPU and 4-GPU Tests (MI30x + MI35x)
       - nightly-accuracy-2-gpu-rocm720
       - nightly-accuracy-2-gpu-vlm-rocm720

test/registered/amd/accuracy/test_qwen38_flash_next_fp8_eval.py

@@ -0,0 +1,228 @@
+"""AMD Qwen3.8-Flash-Next-FP8 graph-mode GSM8K accuracy test.
+
+The released FP8 checkpoint uses TP1 on gfx950 and TP2+EP2 on gfx942, where
+the 192 GiB device capacity cannot hold the target, MTP draft, and graph/KV
+state on one GPU. These nightly tests pin the checkpoint revision and exercise
+the same AITER decode-graph path with EAGLE speculation on both architectures.
+Direct AITER paged QSA remains disabled here so the core model correctness gate
+does not depend on an unreleased AITER API.
+
+Registries:
+  nightly-amd-2-gpu-mi30x-qwen38-flash-next-fp8 suite
+  nightly-amd-1-gpu-mi35x-qwen38-flash-next-fp8 suite
+"""
+
+import base64
+import io
+import os
+import unittest
+from pathlib import Path
+from types import SimpleNamespace
+
+import requests
+
+from sglang.test.ci.ci_register import register_amd_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    terminate_and_kill_process_tree,
+    write_github_step_summary,
+)
+
+register_amd_ci(
+    est_time=1800,
+    suite="nightly-amd-2-gpu-mi30x-qwen38-flash-next-fp8",
+    nightly=True,
+)
+register_amd_ci(
+    est_time=1800,
+    suite="nightly-amd-1-gpu-mi35x-qwen38-flash-next-fp8",
+    nightly=True,
+)
+
+MODEL_ID = "Qwen/Qwen3.8-Flash-Next-FP8"
+MODEL_PATH = os.environ.get("QWEN38_FLASH_NEXT_FP8_MODEL_PATH", MODEL_ID)
+MODEL_REVISION = "bcd9f01ddc9cff2316eb84281bebcd5b058bddce"
+ACCURACY_THRESHOLD = 0.94
+SERVER_LAUNCH_TIMEOUT = 1800
+NUM_REQUESTED_EXAMPLES = 1319
+NUM_SHOTS = 5
+NUM_EVALUATED_EXAMPLES = NUM_REQUESTED_EXAMPLES - NUM_SHOTS
+IMAGE_PATH = Path(__file__).resolve().parents[4] / "examples/assets/example_image.png"
+TP_SIZE = int(os.environ.get("QWEN38_FLASH_NEXT_FP8_TP_SIZE", "1"))
+EP_SIZE = int(os.environ.get("QWEN38_FLASH_NEXT_FP8_EP_SIZE", "1"))
+MEM_FRACTION_STATIC = os.environ.get(
+    "QWEN38_FLASH_NEXT_FP8_MEM_FRACTION_STATIC", "0.95"
+)
+
+if TP_SIZE % EP_SIZE != 0:
+    raise ValueError(
+        f"Qwen3.8 nightly requires TP ({TP_SIZE}) divisible by EP ({EP_SIZE})"
+    )
+
+SERVER_ARGS = [
+    "--revision",
+    MODEL_REVISION,
+    "--served-model-name",
+    MODEL_ID,
+    "--tp-size",
+    str(TP_SIZE),
+    "--ep-size",
+    str(EP_SIZE),
+    "--attention-backend",
+    "aiter",
+    "--moe-runner-backend",
+    "aiter",
+    "--kv-cache-dtype",
+    "auto",
+    "--chunked-prefill-size",
+    "16384",
+    "--watchdog-timeout",
+    "1200",
+    "--mem-fraction-static",
+    MEM_FRACTION_STATIC,
+    "--max-running-requests",
+    "4",
+    "--cuda-graph-max-bs-decode",
+    "4",
+    "--speculative-algorithm",
+    "EAGLE",
+    "--speculative-num-steps",
+    "3",
+    "--speculative-eagle-topk",
+    "1",
+    "--speculative-num-draft-tokens",
+    "4",
+]
+
+SERVER_ENV = {
+    "SGLANG_USE_AITER": "0",
+}
+
+
+class TestQwen38FlashNextFP8AMD(CustomTestCase):
+    """Gate Qwen3.8-Flash-Next FP8 accuracy on the graph fallback path."""
+
+    def _assert_multimodal_generation(self):
+        image_data = base64.b64encode(IMAGE_PATH.read_bytes()).decode("ascii")
+        response = requests.post(
+            DEFAULT_URL_FOR_TEST + "/v1/chat/completions",
+            json={
+                "model": MODEL_ID,
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "image_url",
+                                "image_url": {
+                                    "url": f"data:image/png;base64,{image_data}"
+                                },
+                            },
+                            {
+                                "type": "text",
+                                "text": "What color are the taxis? Answer with one word.",
+                            },
+                        ],
+                    }
+                ],
+                "temperature": 0,
+                "max_tokens": 32,
+                "chat_template_kwargs": {"enable_thinking": False},
+            },
+            timeout=120,
+        )
+        response.raise_for_status()
+        content = response.json()["choices"][0]["message"]["content"]
+        self.assertIn("yellow", content.lower())
+
+    def test_gsm8k_accuracy(self):
+        server_stdout = io.StringIO()
+        server_stderr = io.StringIO()
+        process = popen_launch_server(
+            MODEL_PATH,
+            DEFAULT_URL_FOR_TEST,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=SERVER_ARGS,
+            env=SERVER_ENV,
+            return_stdout_stderr=(server_stdout, server_stderr),
+        )
+
+        try:
+            self._assert_multimodal_generation()
+            requests.post(
+                DEFAULT_URL_FOR_TEST + "/flush_cache",
+                params={"timeout": 30},
+                timeout=45,
+            ).raise_for_status()
+            args = SimpleNamespace(
+                base_url=DEFAULT_URL_FOR_TEST,
+                model=MODEL_ID,
+                eval_name="gsm8k",
+                num_examples=NUM_REQUESTED_EXAMPLES,
+                num_threads=512,
+                num_shots=NUM_SHOTS,
+                max_tokens=4096,
+                chat_template_kwargs={"enable_thinking": False},
+            )
+            metrics = run_eval(args)
+            score = metrics["score"]
+            latency = metrics.get("latency", 0.0)
+            output_throughput = metrics.get("output_throughput", 0.0)
+            status = "PASS" if score >= ACCURACY_THRESHOLD else "FAIL"
+
+            self.assertIsNone(
+                process.poll(), "Qwen3.8-Flash-Next server exited during evaluation"
+            )
+            server_logs = server_stdout.getvalue() + server_stderr.getvalue()
+            decode_lines = [
+                line for line in server_logs.splitlines() if "Decode batch" in line
+            ]
+            self.assertTrue(
+                decode_lines,
+                "Qwen3.8-Flash-Next completed without decode batch evidence",
+            )
+            eager_decode_lines = [
+                line for line in decode_lines if "cuda graph: True" not in line
+            ]
+            self.assertFalse(
+                eager_decode_lines,
+                "Qwen3.8-Flash-Next used eager decode instead of graph replay: "
+                + "\n".join(eager_decode_lines[:5]),
+            )
+            self.assertNotRegex(
+                server_logs,
+                r"Parameter .*not found in params_dict",
+                "Qwen3.8-Flash-Next skipped checkpoint parameters during loading",
+            )
+
+            if is_in_ci():
+                summary = "### Qwen3.8-Flash-Next-FP8 GSM8K (AMD)\n\n"
+                summary += (
+                    "| Model revision | TP | EP | Graph | Responses | Accuracy | "
+                    "Threshold | Latency | Output throughput | Status |\n"
+                )
+                summary += "| --- | ---: | ---: | --- | ---: | ---: | ---: | ---: | ---: | --- |\n"
+                summary += (
+                    f"| `{MODEL_REVISION}` | {TP_SIZE} | {EP_SIZE} | yes | "
+                    f"{NUM_EVALUATED_EXAMPLES} | "
+                    f"{score:.3f} | {ACCURACY_THRESHOLD:.2f} | {latency:.1f}s | "
+                    f"{output_throughput:.1f} tok/s | {status} |\n"
+                )
+                write_github_step_summary(summary)
+
+            self.assertGreaterEqual(
+                score,
+                ACCURACY_THRESHOLD,
+                f"Qwen3.8-Flash-Next-FP8 accuracy {score:.3f} below "
+                f"threshold {ACCURACY_THRESHOLD:.2f}",
+            )
+        finally:
+            terminate_and_kill_process_tree(process)
+
+
+if __name__ == "__main__":
+    unittest.main()