[AMD][DI][CI] 2/N Add DSV4 DP8/EP8 and MTP MI355X 1P1D nightly recipes (#29784)

Co-authored-by: bingxche <bingxche@amd.com>
2026-07-01 23:46:48 +08:00
parent 2a6e5c60fe
commit 9bb7de9258
15 changed files with 970 additions and 3 deletions
@@ -142,8 +142,40 @@ jobs:
          # Manual dispatch input wins; otherwise use the latest image resolved
          # by the setup job; otherwise the launcher falls back to the recipe default.
          IMAGE_OVERRIDE: ${{ inputs.image != '' && inputs.image || needs.setup.outputs.image }}
+          # Keep the scheduler off mia1-p01-g20: its ionic RDMA driver ABI
+          # mismatches the container, so MORI reports "no active RDMA device"
+          # and the disagg server dies on init. Remove once the node is fixed.
+          SLURM_EXCLUDE: mia1-p01-g20
        run: bash scripts/ci/slurm/launch_mi355x.sh

+      - name: Pack logs
+        if: always()
+        run: |
+          LOG_DIR="$HOME/.mi355x_ci/${MATRIX_CONFIG_NAME}"
+          LOG_BUNDLE="${GITHUB_WORKSPACE}/${RESULT_FILENAME}_logs.tar.gz"
+          if [ ! -d "$LOG_DIR" ]; then
+            echo "WARN: log directory not found: $LOG_DIR"
+            exit 0
+          fi
+
+          mapfile -d '' LOG_FILES < <(
+            find "$LOG_DIR" -maxdepth 1 -type f \
+              \( -name '*.log' -o -name 'server_exit_*' -o -name 'bench_exit' \) \
+              -printf '%P\0' | sort -z
+          )
+          if [ "${#LOG_FILES[@]}" -eq 0 ]; then
+            echo "WARN: no log files found in $LOG_DIR"
+            exit 0
+          fi
+
+          STAGING_DIR="$(mktemp -d)"
+          trap 'rm -rf "$STAGING_DIR"' EXIT
+          for log_file in "${LOG_FILES[@]}"; do
+            cp -a "$LOG_DIR/$log_file" "$STAGING_DIR/$log_file" || true
+          done
+          tar czf "$LOG_BUNDLE" -C "$STAGING_DIR" .
+          echo "Packed ${#LOG_FILES[@]} log file(s) -> $LOG_BUNDLE"
+
      - name: Process results
        if: always()
        run: |
@@ -164,7 +196,9 @@ jobs:
        uses: actions/upload-artifact@v4
        with:
          name: mi355x-${{ matrix.config.name }}-${{ github.run_id }}
-          path: ${{ github.workspace }}/*.json
+          path: |
+            ${{ github.workspace }}/*.json
+            ${{ github.workspace }}/*_logs.tar.gz
          retention-days: 30
          if-no-files-found: warn

@@ -22,6 +22,8 @@
 #   MODEL_PATH         - local snapshot dir (preferred over downloading MODEL)
 #   SLURM_PARTITION    - default: amd-sglang
 #   SLURM_NODELIST     - optional explicit node pin (else scheduler chooses)
+#   SLURM_EXCLUDE      - optional comma-separated nodes to keep the scheduler
+#                        off (e.g. hosts with a broken RDMA driver)
 #   RUNNER_NAME        - GitHub runner name (a built-in default env var)
 #   GITHUB_RUN_ID      - GitHub Actions run id (a built-in default env var)
 #                        The allocation is named
@@ -101,6 +103,13 @@ emit("CHUNK", rt["chunked_prefill_size"])
 emit("SWA", rt["swa_full_tokens_ratio"])
 emit("PTP", b["prefill"]["tensor-parallel-size"])
 emit("DTP", b["decode"]["tensor-parallel-size"])
+emit("PEP", b["prefill"].get("expert-parallel-size", 1))
+emit("PDP", b["prefill"].get("data-parallel-size", 1))
+m = r.get("mtp", {}) or {}
+emit("MTP_ENABLED", 1 if m.get("enabled") else 0)
+emit("MTP_STEPS", m.get("num_steps", 3))
+emit("MTP_TOPK", m.get("eagle_topk", 1))
+emit("MTP_DRAFT", m.get("num_draft_tokens", 4))
 # Worker counts double as node counts here: one server per node (TP == GPUs/node).
 # 1P1D today; bumping these reserves 2P2D / 1P3D / 3P1D. Multi-node-per-worker
 # (TP > GPUs/node, needs --dist-init-addr/--nnodes/--node-rank) is out of scope.
@@ -175,12 +184,24 @@ DSV4_ENV=(
 DSV4_ENV_STR="${DSV4_ENV[*]}"
 MORI_ENV="-e MORI_DISABLE_AUTO_XGMI=1 -e NCCL_IB_HCA=ionic -e NCCL_IB_GID_INDEX=1 -e NCCL_CROSS_NIC=1"

+# Optional topology / speculative-decode flags driven by the recipe. Base recipes
+# (EP1/DP1, no mtp) leave EXTRA_FLAGS empty, preserving prior behavior exactly.
+EXTRA_FLAGS=""
+(( PDP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --enable-dp-attention --dp-size $PDP"
+(( PEP > 1 )) && EXTRA_FLAGS="$EXTRA_FLAGS --ep-size $PEP"
+if [[ "$MTP_ENABLED" == "1" ]]; then
+    EXTRA_FLAGS="$EXTRA_FLAGS --speculative-algorithm EAGLE \
+--speculative-num-steps $MTP_STEPS --speculative-eagle-topk $MTP_TOPK \
+--speculative-num-draft-tokens $MTP_DRAFT"
+fi
+echo "extra flags: ${EXTRA_FLAGS:-<none>} (pep=$PEP pdp=$PDP mtp=$MTP_ENABLED)"
+
 COMMON_FLAGS="--trust-remote-code --tp $PTP --disable-radix-cache \
 --attention-backend $ATTN --max-running-requests $MAXREQ --page-size $PAGE \
 --mem-fraction-static $MEMFRAC --swa-full-tokens-ratio $SWA \
 --chunked-prefill-size $CHUNK --disable-shared-experts-fusion \
 --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4 \
--disaggregation-transfer-backend mori --disaggregation-ib-device $IB"
+--disaggregation-transfer-backend mori --disaggregation-ib-device $IB$EXTRA_FLAGS"

 DOCKER_COMMON="--rm --network host --ipc host --shm-size 32g --privileged \
 --security-opt seccomp=unconfined \
@@ -367,6 +388,12 @@ NODELIST_ARG=()
 EXCLUSIVE_ARG=()
 [[ "${SLURM_EXCLUSIVE:-1}" == "1" ]] && EXCLUSIVE_ARG=(--exclusive)

+# Keep the scheduler off known-bad nodes (e.g. a host whose ionic RDMA driver
+# ABI mismatches the container, where MORI reports "no active RDMA device" and
+# the disagg server dies on init). Comma-separated node list.
+EXCLUDE_ARG=()
+[[ -n "${SLURM_EXCLUDE:-}" ]] && EXCLUDE_ARG=(--exclude="$SLURM_EXCLUDE")
+
 # One node per prefill/decode worker (TP == GPUs/node). 1P1D -> 2 nodes.
 TOTAL_NODES=$((PW + DW))

@@ -377,7 +404,7 @@ TOTAL_NODES=$((PW + DW))
 JOB_NAME="mi355x-ci-${RUNNER_NAME:-norunner}-${GITHUB_RUN_ID:-0}-${MATRIX_CONFIG_NAME}"

 set +e
-salloc -p "$SLURM_PARTITION" -N"$TOTAL_NODES" "${NODELIST_ARG[@]}" "${EXCLUSIVE_ARG[@]}" \
+salloc -p "$SLURM_PARTITION" -N"$TOTAL_NODES" "${NODELIST_ARG[@]}" "${EXCLUDE_ARG[@]}" "${EXCLUSIVE_ARG[@]}" \
    --job-name "$JOB_NAME" -t "$TIME_LIMIT" \
    bash "$WORKDIR/drive.sh" "$WORKDIR" "$PW" "$DW"
 SALLOC_RC=$?
@@ -117,3 +117,206 @@ dsv4pro-fp4-mi355x-sglang:
      search-space:
        - conc-list: [1, 8, 16, 32, 64, 128, 256]
          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d.yaml
+
+# AMD MI355X 2-node 1P1D disaggregation over MORI with extra topology / MTP
+# coverage on the SAME four DeepSeek-V4 model x precision combos as the base
+# MORI blocks above (which stay TP8, no MTP). Three variants per model:
+#   * -mtp        : TP8 + EAGLE MTP (recipe `mtp.enabled`)
+#   * -dp8ep8     : DP-attention 8 + narrow within-node EP8
+#   * -dp8ep8-mtp : DP8 + narrow EP8 + EAGLE MTP
+# launch_mi355x.sh reads expert-/data-parallel-size and the `mtp:` section to
+# append --ep-size / --enable-dp-attention --dp-size / --speculative-* flags.
+
+dsv4flash-fp8-mi355x-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-mtp.yaml
+
+dsv4flash-fp8-mi355x-dp8ep8-sglang:
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-dp8ep8.yaml
+
+dsv4flash-fp8-mi355x-dp8ep8-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Flash-FP8
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Flash-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4flash/1k1k/1p1d-dp8ep8-mtp.yaml
+
+dsv4pro-fp8-mi355x-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Pro-FP8
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        # conc256 excluded: disagg-decode SWA hybrid pool retract->get_cpu_copy
+        # is an upstream NotImplementedError (crashes decode). See recipe.
+        - conc-list: [1, 8, 16, 32, 64, 128]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-mtp.yaml
+
+dsv4pro-fp8-mi355x-dp8ep8-sglang:
+  model: sgl-project/DeepSeek-V4-Pro-FP8
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-dp8ep8.yaml
+
+dsv4pro-fp8-mi355x-dp8ep8-mtp-sglang:
+  model: sgl-project/DeepSeek-V4-Pro-FP8
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--sgl-project--DeepSeek-V4-Pro-FP8
+  runner: mi355x
+  precision: fp8
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp8/dsv4pro/1k1k/1p1d-dp8ep8-mtp.yaml
+
+dsv4flash-fp4-mi355x-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mtp.yaml
+
+dsv4flash-fp4-mi355x-dp8ep8-sglang:
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-dp8ep8.yaml
+
+dsv4flash-fp4-mi355x-dp8ep8-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Flash
+  model-prefix: dsv4flash
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-dp8ep8-mtp.yaml
+
+dsv4pro-fp4-mi355x-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-mtp.yaml
+
+dsv4pro-fp4-mi355x-dp8ep8-sglang:
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-dp8ep8.yaml
+
+dsv4pro-fp4-mi355x-dp8ep8-mtp-sglang:
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4pro
+  model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Pro
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: true
+  disagg: true
+  seq-len-configs:
+    - isl: 1024
+      osl: 1024
+      search-space:
+        - conc-list: [1, 8, 16, 32, 64, 128, 256]
+          config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4pro/1k1k/1p1d-dp8ep8-mtp.yaml
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash FP4 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Flash FP4 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash FP4 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Pro FP4 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Pro FP4 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Pro FP4 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Flash-FP8 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,61 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,52 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — DP8 + narrow EP8.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 8
+      data-parallel-size: 8
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  concurrencies: [1, 8, 16, 32, 64, 128, 256]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91
@@ -0,0 +1,68 @@
+# MI355X DeepSeek-V4-Pro-FP8 2-node 1P1D disaggregation recipe — TP8 + MTP.
+#
+# Consumed by:
+#   * scripts/ci/slurm/process_result.py reads `resources` and
+#     `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table.
+#   * scripts/ci/slurm/launch_mi355x.sh reads `runtime`, `bench`, and `mtp`.
+
+resources:
+  prefill_workers: 1
+  decode_workers: 1
+
+backend:
+  sglang_config:
+    prefill:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+    decode:
+      tensor-parallel-size: 8
+      expert-parallel-size: 1
+      data-parallel-size: 1
+
+runtime:
+  image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623
+  attention_backend: dsv4
+  # RoCE HCAs MORI uses for cross-node KV transfer.
+  ib_devices: rdma0,rdma1,rdma2,rdma3
+  prefill_port: 30025
+  decode_port: 30026
+  prefill_bootstrap_port: 8998
+  decode_bootstrap_port: 9001
+  lb_port: 8000
+  mem_fraction_static: 0.90
+  page_size: 256
+  max_running_requests: 256
+  chunked_prefill_size: 8192
+  swa_full_tokens_ratio: 0.1
+
+# MTP / EAGLE speculative decoding (NextN head from the base model). Applied to
+# both prefill and decode. Flags mirror
+# test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py.
+mtp:
+  enabled: true
+  num_steps: 3
+  eagle_topk: 1
+  num_draft_tokens: 4
+
+bench:
+  # bench_serving --max-concurrency sweep; one result JSON per concurrency.
+  # conc256 is excluded for this leg only: at conc256 the disagg-decode SWA
+  # hybrid KV pool fills, the scheduler retracts running requests, and the
+  # retract->offload_kv_cache path calls get_cpu_copy() which is unimplemented
+  # for the SWA hybrid pool (raises NotImplementedError, crashes the decode
+  # scheduler). Raising swa_full_tokens_ratio (tried up to 0.3) does not help
+  # -- usage climbs to fill the larger budget and still retracts. Until the
+  # upstream get_cpu_copy stub is implemented, cap this leg at 128.
+  concurrencies: [1, 8, 16, 32, 64, 128]
+  num_prompts_factor: 4      # num-prompts = concurrency * factor
+  random_range_ratio: 1.0
+
+  # Correctness gate run through the PD path before the perf sweep (full GSM8K,
+  # 8-shot, accuracy > 0.91). A regression here fails the nightly even when
+  # throughput looks fine ("fast but wrong").
+  accuracy:
+    enabled: true
+    num_shots: 8
+    num_questions: 1319    # full GSM8K test set
+    threshold: 0.91