[AMD][CI] Add DeepSeek-V4-Flash FP8 accuracy coverage on MI30x (#36396)

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

@@ -59,11 +59,9 @@ on:
           # 8-GPU Grok2 (MI30x + MI35x)
           - nightly-8-gpu-grok2-rocm720
           - nightly-8-gpu-mi35x-grok2-rocm720
-          # 8-GPU DeepSeek-V3.x (MI30x)
+          # 8-GPU DeepSeek-V3.2 (MI30x)
-          - nightly-8-gpu-deepseek-v31-rocm720
           - nightly-8-gpu-deepseek-v32-rocm720
           - nightly-8-gpu-deepseek-v32-mtp-rocm720
-          - nightly-8-gpu-deepseek-v3-kv-fp8-rocm720
           # 8-GPU DeepSeek-V3.2 (MI35x)
           - nightly-accuracy-8-gpu-mi35x-deepseek-v32-rocm720
           - nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp-rocm720
@@ -75,7 +73,9 @@ on:
           - 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)
+          # 8-GPU DeepSeek-V4-Flash (MI30x)
+          - nightly-8-gpu-deepseek-v4-flash-rocm720
+          # 8-GPU DeepSeek-V4 (MI35x)
           - nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720
           - nightly-8-gpu-mi35x-deepseek-v4-pro-rocm720
           - nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720
@@ -856,60 +856,13 @@ jobs:
           exit ${TEST_EXIT_CODE:-0}
 
   # ==============================================================================
-  # 8-GPU DeepSeek-V3.x (MI30x)
+  # 8-GPU DeepSeek-V3.2 (MI30x)
+  #
+  # V3-0324 and V3.1 are no longer scheduled here: V3.2 covers the same MI30x
+  # paths (aiter MLA, EAGLE MTP, multithread weight load) on a current
+  # checkpoint, and DeepSeek-R1/V4 carry the MI35x side.
   # ==============================================================================
 
-  nightly-8-gpu-deepseek-v31-rocm720:
-    name: ${{ format('nightly-8-gpu-deepseek-v31 ({0}, linux-mi300-8gpu-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-8-gpu-deepseek-v31-rocm720,'))
-    runs-on: linux-mi300-8gpu-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 --skip-test-time-deps
-
-      - name: Accuracy Test ROCm 7.2 (8-GPU DeepSeek-V3.1)
-        timeout-minutes: 120
-        run: |
-          > github_summary.md  # Clear summary file
-          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-            -e SGLANG_USE_AITER=1 \
-            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v31 --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}
-
-      - name: Performance Test ROCm 7.2 (8-GPU DeepSeek-V3.1)
-        timeout-minutes: 300
-        continue-on-error: true
-        run: |
-          > github_summary.md  # Clear summary file
-          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-            -e SGLANG_USE_ROCM700A=1 \
-            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
-            python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-deepseek-v31 --nightly --timeout-per-file 18000 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
-          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
-          exit ${TEST_EXIT_CODE:-0}
-
   nightly-8-gpu-deepseek-v32-rocm720:
     name: ${{ format('nightly-8-gpu-deepseek-v32 ({0}, linux-mi300-8gpu-sglang)', matrix.rocm_version) }}
     strategy:
@@ -1008,44 +961,6 @@ jobs:
           echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
           exit ${TEST_EXIT_CODE:-0}
 
-  nightly-8-gpu-deepseek-v3-kv-fp8-rocm720:
-    name: ${{ format('nightly-8-gpu-deepseek-v3-kv-fp8 ({0}, linux-mi300-8gpu-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-8-gpu-deepseek-v3-kv-fp8-rocm720,'))
-    runs-on: linux-mi300-8gpu-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 --skip-test-time-deps
-
-      - name: DeepSeek-V3 KV FP8 Test ROCm 7.2 (8-GPU Basic + MTP)
-        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-8-gpu-deepseek-v3-kv-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}
-
   # ==============================================================================
   # 8-GPU DeepSeek-V3.2 (MI35x)
   # ==============================================================================
@@ -1465,7 +1380,54 @@ jobs:
           exit ${TEST_EXIT_CODE:-0}
 
   # ==============================================================================
-  # 8-GPU DeepSeek-V4 (MI35x only)
+  # 8-GPU DeepSeek-V4-Flash (MI30x)
+  #
+  # Accuracy only: the MI35x jobs below carry the DSV4 perf numbers, and gfx942
+  # has no DSV4 perf baseline to regress against yet. Runs the cookbook's
+  # MI300X Flash FP8 low-latency cell, so it also covers MLA + KV-FP8 on gfx942.
+  # ==============================================================================
+
+  nightly-8-gpu-deepseek-v4-flash-rocm720:
+    name: ${{ format('nightly-8-gpu-deepseek-v4-flash ({0}, linux-mi300-8gpu-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-8-gpu-deepseek-v4-flash-rocm720,'))
+    runs-on: linux-mi300-8gpu-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 --skip-test-time-deps
+
+      - name: Accuracy Test ROCm 7.2 (8-GPU DeepSeek-V4-Flash FP8)
+        timeout-minutes: 180
+        run: |
+          > github_summary.md  # Clear summary file
+          bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
+            -e SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D=1 \
+            -e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
+            python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v4-flash --nightly --timeout-per-file 5400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
+          echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
+          exit ${TEST_EXIT_CODE:-0}
+
+  # ==============================================================================
+  # 8-GPU DeepSeek-V4 (MI35x)
   # ==============================================================================
 
   nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720:
@@ -2392,11 +2354,9 @@ jobs:
       # 8-GPU Grok2 (MI30x + MI35x)
       - nightly-8-gpu-grok2-rocm720
       - nightly-8-gpu-mi35x-grok2-rocm720
-      # 8-GPU DeepSeek-V3.x (MI30x)
+      # 8-GPU DeepSeek-V3.2 (MI30x)
-      - nightly-8-gpu-deepseek-v31-rocm720
       - nightly-8-gpu-deepseek-v32-rocm720
       - nightly-8-gpu-deepseek-v32-mtp-rocm720
-      - nightly-8-gpu-deepseek-v3-kv-fp8-rocm720
       # 8-GPU DeepSeek-V3.2 (MI35x)
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32-rocm720
       - nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp-rocm720
@@ -2408,7 +2368,9 @@ jobs:
       - 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)
+      # 8-GPU DeepSeek-V4-Flash (MI30x)
+      - nightly-8-gpu-deepseek-v4-flash-rocm720
+      # 8-GPU DeepSeek-V4 (MI35x)
       - nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720
       - nightly-8-gpu-mi35x-deepseek-v4-pro-rocm720
       - nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720

test/registered/amd/test_deepseek_v4_flash_fp8_mi30x.py

@@ -0,0 +1,139 @@
+"""MI30x DeepSeek-V4-Flash FP8 Accuracy Test (8-GPU)
+
+GSM8K few-shot accuracy for DeepSeek-V4-Flash FP8 on MI30x (gfx942) ROCm 7.2.
+
+The launch config is the cookbook's MI300X Flash FP8 low-latency single-node
+recipe, verbatim -- the one the docs tell users to run, marked `verified: true`
+in docs/src/snippets/configs/deepseek-ai/deepseek-v4.jsx. Of the three published
+MI300X strategies it is the only TP-only one (balanced and high-throughput add
+DP attention and the prefill delayer), so it is the narrowest cell that still
+covers the gfx942 serving path. Keep this in sync with that cell.
+
+The cell as published could not load this checkpoint; the fix is in the
+cookbook PR and is the SGLANG_DSV4_FP4_EXPERTS entry below.
+
+Accuracy only: gfx942 has no DSV4 perf baseline to regress against yet, and the
+MI35x suite already carries the 8k/1k throughput numbers.
+
+Unlike the MI35x FP8 test this recipe also runs `--kv-cache-dtype fp8_e4m3` and
+EAGLE, so it is the gfx942 MLA + KV-FP8 signal as well. That extra quantization
+is also why the threshold below is 0.90 rather than the 0.91 the MI35x FP8 test
+uses: measured 0.9174 and 0.9257 here, a spread wide enough that 0.91 would sit
+only ~2 sd off the mean and fail a good build roughly one night in forty. 0.90
+keeps ~2 points of headroom and still catches every regression worth gating on
+-- the failure modes this guards against (see #36390) collapse output entirely
+rather than shaving a point.
+
+Registry: nightly-amd-accuracy-8-gpu-deepseek-v4-flash suite
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+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_amd_ci(
+    est_time=5400,
+    suite="nightly-amd-accuracy-8-gpu-deepseek-v4-flash",
+    nightly=True,
+)
+
+DEEPSEEK_V4_FP8_MODEL_PATH = os.environ.get(
+    "DEEPSEEK_V4_FP8_MODEL_PATH", "sgl-project/DeepSeek-V4-Flash-FP8"
+)
+SERVER_LAUNCH_TIMEOUT = 3600
+
+ENV_VARS = {
+    # The routed experts of this checkpoint are FP8, not mxfp4-packed. The env
+    # default is mxfp4 and the auto-detect fallback reads the safetensors header
+    # from the local HF cache, so it returns None on a runner that has not pulled
+    # the weights yet and the wrong default wins -- weight load then dies on a
+    # factor-of-2 shape mismatch. Set it rather than depend on cache state.
+    "SGLANG_DSV4_FP4_EXPERTS": "0",
+    "SGLANG_USE_ROCM700A": "0",
+    "SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
+    "AITER_BF16_FP8_MOE_BOUND": "0",
+}
+
+
+class TestDeepseekV4FlashFp8Mi30x(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEEPSEEK_V4_FP8_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+
+        env = os.environ.copy()
+        env.update(ENV_VARS)
+
+        other_args = [
+            "--trust-remote-code",
+            "--tp",
+            "8",
+            "--attention-backend",
+            "dsv4",
+            "--page-size",
+            "256",
+            "--mem-fraction-static",
+            "0.90",
+            "--swa-full-tokens-ratio",
+            "0.1",
+            "--disable-shared-experts-fusion",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--chunked-prefill-size",
+            "8192",
+            "--speculative-algorithm",
+            "EAGLE",
+            "--speculative-num-steps",
+            "3",
+            "--speculative-eagle-topk",
+            "1",
+            "--speculative-num-draft-tokens",
+            "4",
+        ]
+
+        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_gsm8k(self):
+        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-v4-flash-fp8, mi30x)\n"
+                f'{metrics["accuracy"]=:.3f}\n'
+            )
+            self.assertGreater(metrics["accuracy"], 0.90)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/run_suite.py

@@ -147,6 +147,7 @@ NIGHTLY_SUITES = {
         "nightly-amd-4-gpu",
         "nightly-amd-8-gpu",
         "nightly-amd-vlm",
+        "nightly-amd-accuracy-8-gpu-deepseek-v4-flash",
         "nightly-amd-8-gpu-mi35x-deepseek-v4-flash",
         # MI35x 8-GPU suite (different model configs)
         "nightly-amd-8-gpu-mi35x",