diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 74dce63495..fb219570fc 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -19,6 +19,115 @@ check_env_vars() {
     fi
 }
 
+# Report live members of explicitly owned process groups. Zombies cannot hold
+# output pipes open. Do not use leader liveness: a router can orphan its workers.
+_background_process_groups_alive() {
+    local groups=" $* "
+    local listing
+    listing=$(ps -eo pgid=,stat=) || return 1
+    awk -v groups="$groups" '
+        index(groups, " " $1 " ") && $2 !~ /^[ZX]/ { alive[$1] = 1 }
+        END { for (group in alive) print group }
+    ' <<< "$listing"
+}
+
+# Called only after benchmark/eval work ends. Preserve its exit status while
+# bounding teardown of the setsid groups recorded by the launcher. Grace periods
+# are explicit arguments, independent of benchmark duration and server readiness.
+stop_background_process_groups() {
+    local work_status="$1" term_grace="$2" kill_grace="$3"
+    shift 3
+    local pgid own_pgid remaining deadline cleanup_status=0
+    local groups=("$@")
+    if [[ ! "$work_status" =~ ^[0-9]+$ || ! "$term_grace" =~ ^[0-9]+$ || ! "$kill_grace" =~ ^[0-9]+$ ]]; then
+        echo "ERROR: invalid process-group cleanup status or grace period" >&2
+        return 1
+    fi
+    if ! own_pgid=$(ps -o pgid= -p "$$"); then
+        if [[ "$work_status" -ne 0 ]]; then return "$work_status"; fi
+        return 1
+    fi
+    own_pgid="${own_pgid//[[:space:]]/}"
+    for pgid in "${groups[@]}"; do
+        if [[ ! "$pgid" =~ ^[1-9][0-9]*$ || "$pgid" -le 1 || "$pgid" == "$own_pgid" ]]; then
+            echo "ERROR: refusing unsafe process-group cleanup: '$pgid'" >&2
+            if [[ "$work_status" -ne 0 ]]; then return "$work_status"; fi
+            return 1
+        fi
+    done
+    if [[ ${#groups[@]} -eq 0 ]]; then return "$work_status"; fi
+
+    echo "Stopping owned process groups: ${groups[*]}"
+    for pgid in "${groups[@]}"; do
+        kill -TERM -- "-$pgid" 2>/dev/null || true
+    done
+    deadline=$((SECONDS + term_grace))
+    while true; do
+        remaining=$(_background_process_groups_alive "${groups[@]}") || { cleanup_status=1; break; }
+        [[ -n "$remaining" && $SECONDS -lt $deadline ]] || break
+        sleep 1
+    done
+    if [[ -n "$remaining" ]]; then
+        echo "TERM grace expired; force-stopping owned process groups: $remaining"
+        for pgid in $remaining; do
+            kill -KILL -- "-$pgid" 2>/dev/null || true
+        done
+        deadline=$((SECONDS + kill_grace))
+        while true; do
+            remaining=$(_background_process_groups_alive "${groups[@]}") || { cleanup_status=1; break; }
+            [[ -n "$remaining" && $SECONDS -lt $deadline ]] || break
+            sleep 1
+        done
+        if [[ -n "$remaining" ]]; then
+            echo "ERROR: process groups still alive after KILL grace: $remaining" >&2
+            cleanup_status=1
+        fi
+    fi
+    if [[ "$work_status" -ne 0 ]]; then return "$work_status"; fi
+    return "$cleanup_status"
+}
+
+# EXIT handler for launchers that own explicit setsid groups and optionally one
+# auxiliary daemon PID. Disable this handler before exiting to avoid recursion.
+# Run auxiliary cleanup even when group cleanup fails, preserving the work code.
+exit_after_background_process_cleanup() {
+    local work_status="$1" term_grace="$2" kill_grace="$3" auxiliary_pid="$4"
+    shift 4
+    local final_status
+    trap - EXIT
+    if stop_background_process_groups "$work_status" "$term_grace" "$kill_grace" "$@"; then
+        final_status=0
+    else
+        final_status=$?
+    fi
+    if [[ -n "$auxiliary_pid" ]]; then
+        kill "$auxiliary_pid" 2>/dev/null || true
+    fi
+    exit "$final_status"
+}
+
+# Finish preflight on every allocated node before any server container starts its
+# peer-readiness deadline. A failed node prevents the entire serving step.
+run_amd_multinode_after_preflight() {
+    local nodelist="$1" node_count="$2" preflight_script="$3"
+    local container_filter="$4" skip_gpu_sanity="$5"
+    shift 5
+    local preflight_rc
+    if srun --nodelist="$nodelist" \
+        --nodes="$node_count" --ntasks="$node_count" --ntasks-per-node=1 \
+        --kill-on-bad-exit=1 --unbuffered \
+        bash "$preflight_script" "$container_filter" "$skip_gpu_sanity"; then
+        echo "[preflight] all nodes ready; launching server containers"
+    else
+        preflight_rc=$?
+        echo "[preflight][ERROR] node preflight failed; no server containers launched" >&2
+        return "$preflight_rc"
+    fi
+    srun --nodelist="$nodelist" \
+        --nodes="$node_count" --ntasks="$node_count" --ntasks-per-node=1 \
+        --kill-on-bad-exit=1 --signal=TERM@30 --unbuffered "$@"
+}
+
 # Launchers may load only input validation, without benchmark initialization.
 if [[ "${1-}" == "--validation-only" ]]; then
     return 0
diff --git a/benchmarks/multi_node/amd_utils/job.slurm b/benchmarks/multi_node/amd_utils/job.slurm
index c12535776c..d1f5ea11a7 100755
--- a/benchmarks/multi_node/amd_utils/job.slurm
+++ b/benchmarks/multi_node/amd_utils/job.slurm
@@ -583,11 +583,10 @@ if [[ -n "${CLIENT_IMAGE:-}" ]]; then
     srun --nodelist="$SELECTED_NODELIST_SRUN" bash -c 'eval "$DOCKER_CMD_DETECT"; $DOCKER_CMD pull '"$CLIENT_IMAGE"' >/dev/null 2>&1 || true' 2>/dev/null || true
 fi
 
-srun \
-  --nodelist="$SELECTED_NODELIST_SRUN" \
-  --kill-on-bad-exit=1 \
-  --signal=TERM@30 \
-  --unbuffered \
+run_amd_multinode_after_preflight \
+  "$SELECTED_NODELIST_SRUN" "$NUM_NODES" \
+  "$DI_REPO_DIR/benchmarks/multi_node/amd_utils/preflight_node.sh" \
+  "$CONT_FILTER" "$SKIP_GPU_SANITY" \
   bash -lc "
 set -eo pipefail
 
@@ -675,23 +674,6 @@ else
 fi
 fi  # end: if ENGINE == atom-disagg
 
-# Pre-clean (idempotent): stop then force-remove so GPU VRAM is released
-# before the drain gate. stop-only left containers in Created/Exited state
-# on some nodes.
-\$DOCKER_CMD ps -aq --filter \"$CONT_FILTER\" | xargs -r \$DOCKER_CMD rm -f || true
-\$DOCKER_CMD ps -aq | xargs -r \$DOCKER_CMD stop -t 15 || true
-\$DOCKER_CMD ps -aq | xargs -r \$DOCKER_CMD rm -f || true
-sleep 2
-
-# GPU drain gate: fail fast on leftover VRAM use instead of OOMing in model
-# load ~15 min later. Reuses wait_for_amd_gpu_clean from benchmark_lib.sh.
-if [[ \"${SKIP_GPU_SANITY}\" == \"1\" ]]; then
-    echo \"[INFO] SKIP_GPU_SANITY=1 set; skipping GPU pre-flight drain check\"
-else
-    # Unset so benchmark_lib.sh's unrelated agentic KV_OFFLOADING check doesn't exit 1 here.
-    bash -c \"unset IS_AGENTIC SCENARIO_TYPE; source $DI_REPO_DIR/benchmarks/benchmark_lib.sh && wait_for_amd_gpu_clean\"
-fi
-
 # Start vLLM external router container on node 0
 if [[ \"$ENGINE\" == \"vllm-disagg\" && \"$ROUTER_TYPE\" == \"vllm-router\" && \"\$SLURM_PROCID\" == \"0\" ]]; then
     \$DOCKER_CMD rm -f \"$ROUTER_CONT_NAME\" 2>/dev/null || true
@@ -775,6 +757,7 @@ echo \"[rank 0] Main container exited (rc=\$DOCKER_EXIT_CODE). Stopping vllm-rou
 \$DOCKER_CMD rm -f \"$ROUTER_CONT_NAME\" 2>/dev/null || true
 exit \$DOCKER_EXIT_CODE
 "
+SERVER_SRUN_RC=$?
 
 if [[ "${KEEP_CONTAINERS}" != "1" ]]; then
     srun --nodelist="$SELECTED_NODELIST_SRUN" bash -c 'eval "$DOCKER_CMD_DETECT"; $DOCKER_CMD rm -f '"$DOCKER_CONT_NAME"' '"$CLIENT_CONT_NAME"' 2>/dev/null || true'
@@ -786,3 +769,29 @@ if [[ "${KEEP_CONTAINERS}" != "1" ]]; then
         '
     fi
 fi
+
+# /run_logs is backed by each compute node's local /tmp, so the node-0 copy
+# performed by the engine launcher cannot see prefill/decode logs written on
+# other nodes. Collect after the server step and container cleanup so failed
+# runs also include shutdown output. KEEP_CONTAINERS=1 retains a snapshot of
+# any containers left running for debugging.
+# Use sudo because the container-created source and the existing node-0
+# destination can be root-owned. Restore ownership after the fan-in so a
+# subsequent runner job can clean the workspace normally.
+SHARED_JOB_LOGS="${BENCHMARK_LOGS_DIR}/logs/slurm_job-${SLURM_JOB_ID}"
+if ! srun --nodelist="$SELECTED_NODELIST_SRUN" \
+    --nodes="$NUM_NODES" --ntasks="$NUM_NODES" --ntasks-per-node=1 \
+    bash "$DI_REPO_DIR/benchmarks/multi_node/amd_utils/stage_node_logs.sh" \
+    "/tmp/slurm_job-${SLURM_JOB_ID}" "$SHARED_JOB_LOGS"; then
+    echo "[logs][ERROR] failed to stage logs from one or more Slurm nodes" >&2
+    if [[ "$SERVER_SRUN_RC" -eq 0 ]]; then
+        SERVER_SRUN_RC=1
+    fi
+fi
+
+if [[ -d "$SHARED_JOB_LOGS" ]]; then
+    sudo chown -R "$(id -u):$(id -g)" "$SHARED_JOB_LOGS" 2>/dev/null || true
+    chmod -R a+rwX "$SHARED_JOB_LOGS" 2>/dev/null || true
+fi
+
+exit "$SERVER_SRUN_RC"
diff --git a/benchmarks/multi_node/amd_utils/models.yaml b/benchmarks/multi_node/amd_utils/models.yaml
index f8507cf0f0..53666ce487 100644
--- a/benchmarks/multi_node/amd_utils/models.yaml
+++ b/benchmarks/multi_node/amd_utils/models.yaml
@@ -354,9 +354,21 @@ DeepSeek-R1-0528-MXFP4-v2:
 
 DeepSeek-V4-Pro-AgentX: &DeepSeek-V4-Pro-AgentX
   base_flags: "--enable-deepseek-v4-fp4-indexer --watchdog-timeout 3600 --load-balance-method round_robin --kv-cache-dtype fp8_e4m3 --attention-backend dsv4 --page-size 256 --swa-full-tokens-ratio 0.1 --enforce-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4 --disaggregation-transfer-backend mori --tokenizer-worker-num 8 --stream-interval 20 --log-level info --log-level-http error"
-  dp_flags: "--enable-dp-attention --swa-full-tokens-ratio 0.15 --enable-dp-attention-local-control-broadcast"
+  # --enable-dp-lm-head is required by SGLang for DSpark under DP attention; it
+  # is harmless for the EAGLE/MTP arms, so it stays unconditional here rather
+  # than needing a second DP flag string.
+  dp_flags: "--enable-dp-attention --enable-dp-lm-head --swa-full-tokens-ratio 0.15 --enable-dp-attention-local-control-broadcast"
   ep_flags: "--ep-dispatch-algorithm fake --moe-a2a-backend mori --deepep-mode normal"
   mtp_flags: "--speculative-algorithm EAGLE --speculative-eagle-topk 1"
+  # DSpark draft head, selected when the sweep sets spec-decoding: draft_model.
+  # Kept alongside mtp_flags rather than replacing it, so recipes that stay on
+  # spec-decoding: mtp keep the EAGLE arm untouched. The -0813 checkpoint
+  # bundles the draft head (dspark_block_size / dspark_markov_rank /
+  # dspark_target_layer_ids in config.json), so --speculative-draft-model-path
+  # defaults to --model-path and no separate draft checkpoint is needed.
+  # server_sglang.sh appends the block size and the verify window from
+  # DECODE_MTP_SIZE (= gamma); unlike EAGLE, num-steps is pinned to 1.
+  dspark_flags: "--speculative-algorithm DSPARK --speculative-eagle-topk 1"
   prefill:
     disable_radix_cache: false
     disable_cuda_graph: true
@@ -384,8 +396,11 @@ DeepSeek-V4-Pro-AgentX: &DeepSeek-V4-Pro-AgentX
       max_running_requests: "BENCH_MAX_CONC_VALUE*2"
       cuda_graph_bs_range: "1-BENCH_MAX_CONC_VALUE*2"
 
-# Pro-0813 retains EAGLE 3-1-4 for PD compatibility. Synthetic acceptance is
-# checkpoint-specific in server_sglang.sh: thinking-on, length 3 uses AL 3.01.
+# Pro-0813 serves the PD path with DSPARK (spec-decoding: draft_model), which
+# measured clean across c4-c256; the EAGLE 3-1-4 arm remains available through
+# spec-decoding: mtp. Synthetic acceptance is checkpoint-specific in
+# server_sglang.sh and does not depend on which of the two runs: thinking-on,
+# draft length 3 uses AL 3.01.
 DeepSeek-V4-Pro-0813-AgentX: *DeepSeek-V4-Pro-AgentX
 
 DeepSeek-V4-Pro-DI:
diff --git a/benchmarks/multi_node/amd_utils/preflight_node.sh b/benchmarks/multi_node/amd_utils/preflight_node.sh
new file mode 100644
index 0000000000..34759bab89
--- /dev/null
+++ b/benchmarks/multi_node/amd_utils/preflight_node.sh
@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/../../benchmark_lib.sh" --validation-only
+check_env_vars DOCKER_CMD_DETECT DI_REPO_DIR SLURM_JOB_ID
+CONT_FILTER="$1"
+SKIP_GPU_SANITY="$2"
+check_env_vars CONT_FILTER SKIP_GPU_SANITY
+
+preflight_node() {
+    eval "$DOCKER_CMD_DETECT"
+
+    # Preserve the existing pre-clean scope and ordering. Moving it into this
+    # separate Slurm step prevents one node starting while another still drains.
+    $DOCKER_CMD ps -aq --filter "$CONT_FILTER" | xargs -r $DOCKER_CMD rm -f || true
+    $DOCKER_CMD ps -aq | xargs -r $DOCKER_CMD stop -t 15 || true
+    $DOCKER_CMD ps -aq | xargs -r $DOCKER_CMD rm -f || true
+    sleep 2
+
+    if [[ "$SKIP_GPU_SANITY" == "1" ]]; then
+        echo "[INFO] SKIP_GPU_SANITY=1 set; skipping GPU pre-flight drain check"
+    else
+        # Avoid benchmark-only agentic initialization on the host, as before.
+        bash -c 'unset IS_AGENTIC SCENARIO_TYPE; source "$DI_REPO_DIR/benchmarks/benchmark_lib.sh" && wait_for_amd_gpu_clean'
+    fi
+}
+
+NODE_LOG_DIR="/tmp/slurm_job-${SLURM_JOB_ID}"
+mkdir -p "$NODE_LOG_DIR"
+preflight_node 2>&1 | tee "$NODE_LOG_DIR/preflight_$(hostname).log"
diff --git a/benchmarks/multi_node/amd_utils/server_sglang.sh b/benchmarks/multi_node/amd_utils/server_sglang.sh
index b44ea1a11f..39ab32be20 100755
--- a/benchmarks/multi_node/amd_utils/server_sglang.sh
+++ b/benchmarks/multi_node/amd_utils/server_sglang.sh
@@ -23,6 +23,12 @@ export BENCH_MAX_CONC_VALUE
 source $SGLANG_WS_PATH/setup_deps.sh
 source $SGLANG_WS_PATH/env.sh
 
+# Install before starting UMBP or serving processes. Early readiness failures must
+# close the same owned groups as normal completion, including orphaned workers.
+SGLANG_OWNED_PGIDS=()
+UMBP_SA_PID=""
+trap 'exit_after_background_process_cleanup "$?" 30 5 "$UMBP_SA_PID" "${SGLANG_OWNED_PGIDS[@]}"' EXIT
+
 host_ip=$(ip route get 1.1.1.1 | awk '/src/ {print $7}')
 host_name=$(hostname)
 
@@ -100,6 +106,7 @@ def parse_range(cuda_range, default_start, default_end):
 # Output shell variables
 print(f'MODEL_BASE_FLAGS=\"{m.get(\"base_flags\", \"\")}\"')
 print(f'MODEL_MTP_FLAGS=\"{m.get(\"mtp_flags\", \"\")}\"')
+print(f'MODEL_DSPARK_FLAGS=\"{m.get(\"dspark_flags\", \"\")}\"')
 print(f'MODEL_DP_FLAGS=\"{m.get(\"dp_flags\", \"\")}\"')
 print(f'MODEL_EP_FLAGS=\"{m.get(\"ep_flags\", \"\")}\"')
 
@@ -381,8 +388,30 @@ build_server_config() {
     local ep_config=""
     local specific_config=""
 
+    # Speculative-decoding config (only if a draft length is set).
+    #
+    # DECODE_MTP_SIZE carries the draft length for BOTH algorithms, but the two
+    # spend it differently:
+    #   EAGLE/MTP   -- num-steps = draft length, i.e. that many sequential draft
+    #                  forward passes, each producing one token.
+    #   DSPARK      -- one draft pass emits a whole block, so num-steps is
+    #                  pinned to 1 and the draft length becomes the block size
+    #                  (gamma). Passing gamma as num-steps here would ask for
+    #                  gamma sequential DSpark passes instead of one gamma-token
+    #                  block.
+    # The verify window (num-draft-tokens = draft length + 1) is the same for
+    # both, which is also what makes the MORI decode dispatch scaling
+    # (x (DECODE_MTP_SIZE + 1)) correct for DSpark without further change.
     if [ "$decode_mtp_size" -gt 0 ]; then
-        mtp_config="${MODEL_MTP_FLAGS} --speculative-num-steps ${decode_mtp_size} --speculative-num-draft-tokens $((decode_mtp_size + 1))"
+        if [[ "${SPEC_DECODING:-}" == "draft_model" ]]; then
+            # MODEL_DSPARK_FLAGS is validated at the call site, not here: this
+            # function is only ever invoked inside $( ), where an exit would
+            # terminate the subshell and leave the caller with an empty config
+            # rather than aborting the launch.
+            mtp_config="${MODEL_DSPARK_FLAGS} --speculative-dspark-block-size ${decode_mtp_size} --speculative-num-steps 1 --speculative-num-draft-tokens $((decode_mtp_size + 1))"
+        else
+            mtp_config="${MODEL_MTP_FLAGS} --speculative-num-steps ${decode_mtp_size} --speculative-num-draft-tokens $((decode_mtp_size + 1))"
+        fi
     fi
 
     if [[ "$enable_dp" == "true" ]]; then
@@ -439,6 +468,15 @@ build_server_config() {
     echo "$full_config"
 }
 
+# Validate the DSpark path before building either config. This has to happen at
+# top level: build_server_config only ever runs inside $( ), so an exit there
+# would kill the subshell and hand the caller an empty config string instead of
+# stopping the launch.
+if [[ "$DECODE_MTP_SIZE" -gt 0 ]] && [[ "${SPEC_DECODING:-}" == "draft_model" ]] && [[ -z "${MODEL_DSPARK_FLAGS// }" ]]; then
+    echo "FATAL: SPEC_DECODING=draft_model but model '${MODEL_NAME}' has no dspark_flags in models.yaml." >&2
+    exit 1
+fi
+
 PREFILL_SERVER_CONFIG=$(build_server_config "prefill" "$MODEL_NAME" "$PREFILL_TP_SIZE" "$PREFILL_ENABLE_EP" "$PREFILL_ENABLE_DP" "$DECODE_MTP_SIZE")
 DECODE_SERVER_CONFIG=$(build_server_config "decode" "$MODEL_NAME" "$DECODE_TP_SIZE" "$DECODE_ENABLE_EP" "$DECODE_ENABLE_DP" "$DECODE_MTP_SIZE")
 
@@ -654,7 +692,6 @@ elif [[ "$KV_OFFLOADING" != "none" && "$KV_OFFLOAD_BACKEND" == umbp-linker* ]];
             "$UMBP_SA_BIN" "$UMBP_STANDALONE_ADDRESS" > "$UMBP_SA_LOG" 2>&1 &
         UMBP_SA_PID=$!
         echo "[UMBP] standalone server PID: $UMBP_SA_PID"
-        trap '[[ -n "${UMBP_SA_PID:-}" ]] && kill "$UMBP_SA_PID" 2>/dev/null || true' EXIT
 
         # Three waits, all bounded by wall time rather than by a guess at how
         # fast this node is. Bind time for a 549 GB tier measured 120 s on
@@ -927,8 +964,8 @@ if [ "$NODE_RANK" -eq 0 ]; then
             > >(tee /run_logs/slurm_job-${SLURM_JOB_ID}/prefill_${host_name}.log >/dev/null) 2>&1 &
         set +x
         prefill0_pid=$!
-        prefill0_pgid=$(ps -o pgid= -p "$prefill0_pid" 2>/dev/null | tr -d ' ')
-        : "${prefill0_pgid:=$prefill0_pid}"
+        prefill0_pgid=$prefill0_pid
+        SGLANG_OWNED_PGIDS+=("$prefill0_pgid")
     fi
 
     echo "Waiting for all prefill and decode servers to be up . . ."
@@ -985,8 +1022,8 @@ if [ "$NODE_RANK" -eq 0 ]; then
         fi
         set +x
         proxy_pid=$!
-        proxy_pgid=$(ps -o pgid= -p "$proxy_pid" 2>/dev/null | tr -d ' ')
-        : "${proxy_pgid:=$proxy_pid}"
+        proxy_pgid=$proxy_pid
+        SGLANG_OWNED_PGIDS+=("$proxy_pgid")
 
         HEALTH_BARRIER_CMD="python3 $SGLANG_WS_PATH/sync.py barrier \
             --node-ips ${NODE0_ADDR} \
@@ -1082,6 +1119,7 @@ if [ "$NODE_RANK" -eq 0 ]; then
         IS_AGENTIC_RUN=1
     fi
 
+    BENCHMARK_EXIT_CODE=0
     if [[ "${EVAL_ONLY}" == "true" ]]; then
         echo "EVAL_ONLY mode: skipping throughput benchmark"
     elif [[ "$DRY_RUN" -eq 1 ]]; then
@@ -1156,10 +1194,12 @@ print(json.dumps(json.loads(sys.stdin.read())))' <<<"$_val")" || {
             --entrypoint "" \
             "${CLIENT_IMAGE}" \
             bash -lc "cd /workspace/benchmarks/multi_node/amd_utils && bash trace_replay.sh /models ${MODEL_NAME} \"${BENCH_MAX_CONCURRENCY}\" /run_logs/slurm_job-${SLURM_JOB_ID}"
+        BENCHMARK_EXIT_CODE=$?
         set +x
     else
         set -x
         eval "$BENCH_CMD"
+        BENCHMARK_EXIT_CODE=$?
         set +x
     fi
 
@@ -1183,6 +1223,8 @@ print(json.dumps(json.loads(sys.stdin.read())))' <<<"$_val")" || {
             # Must run from repo root so infx/evals/gsm8k.yaml resolves
             pushd /workspace
 
+            # Match the disaggregation router launched above.
+            export PORT=30000
             source /workspace/benchmarks/benchmark_lib.sh
 
             # CONC must be exported before run_eval so meta_env.json matches validate_scores.py.
@@ -1204,9 +1246,9 @@ print(json.dumps(json.loads(sys.stdin.read())))' <<<"$_val")" || {
             # arrive via Docker -e flags from job.slurm.
 
             if [[ "$DRY_RUN" -eq 1 ]]; then
-                echo "DRY RUN: run_eval --port 30000 (framework=${EVAL_FRAMEWORK}, conc=${EVAL_CONCURRENT_REQUESTS}, ctx=${EVAL_MAX_MODEL_LEN:-auto})"
+                echo "DRY RUN: run_eval --port ${PORT} (framework=${EVAL_FRAMEWORK}, conc=${EVAL_CONCURRENT_REQUESTS}, ctx=${EVAL_MAX_MODEL_LEN:-auto})"
             else
-                run_eval --port 30000
+                run_eval --port "$PORT"
                 eval_rc=$?
 
                 if [[ $eval_rc -ne 0 ]]; then
@@ -1247,22 +1289,11 @@ print(json.dumps(json.loads(sys.stdin.read())))' <<<"$_val")" || {
         echo "Copied results to $LOGS_OUTPUT/slurm_job-${SLURM_JOB_ID}"
     fi
 
-    echo "Killing the proxy server and prefill server"
-
-    if [[ "$DRY_RUN" -eq 0 ]]; then
-        # Group-kill the router (setsid at launch): the python launcher has usually
-        # exited after spawning the Rust worker, which reparents to init but stays in
-        # this group; kill $proxy_pid alone misses it and :30000 stays open.
-        kill -TERM -"${proxy_pgid:-$proxy_pid}" 2>/dev/null || true
-        # Group-kill the prefill tree so TP-scheduler children release the tee pipe
-        # and the container can exit.
-        kill -TERM -"${prefill0_pgid:-$prefill0_pid}" 2>/dev/null || true
-    fi
-
-    if [[ "${EVAL_FAILED:-0}" -eq 1 ]]; then
-        echo "ERROR: eval failed; exiting node-0 with rc=1"
-        exit 1
+    node_exit_status=$BENCHMARK_EXIT_CODE
+    if [[ "${EVAL_FAILED:-0}" -eq 1 && "$node_exit_status" -eq 0 ]]; then
+        node_exit_status=1
     fi
+    exit "$node_exit_status"
 
 elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$NODE_OFFSET" ]; then
     echo "${host_name}:${host_ip} is Prefill Node (Model: ${MODEL_NAME})"
@@ -1306,8 +1337,8 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$NODE_OFFSET" ]; then
             > >(tee /run_logs/slurm_job-${SLURM_JOB_ID}/prefill_${host_name}.log >/dev/null) 2>&1 &
         set +x
         prefill_pid=$!
-        prefill_pgid=$(ps -o pgid= -p "$prefill_pid" 2>/dev/null | tr -d ' ')
-        : "${prefill_pgid:=$prefill_pid}"
+        prefill_pgid=$prefill_pid
+        SGLANG_OWNED_PGIDS+=("$prefill_pgid")
     fi
 
     echo "Waiting for proxy server to be up..."
@@ -1337,8 +1368,7 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$NODE_OFFSET" ]; then
     echo "Killing the rank $NODE_RANK prefill server"
 
     if [[ "$DRY_RUN" -eq 0 ]]; then
-        # Group-kill so TP-scheduler children release the tee pipe and the container exits.
-        kill -TERM -"${prefill_pgid:-$prefill_pid}" 2>/dev/null || true
+        exit 0
     fi
 
 else
@@ -1393,7 +1423,7 @@ else
             if [[ -n "$DSV4_GOLDEN_AL" ]]; then
                 DECODE_SIM_ACC_ENV="SGLANG_SIMULATE_ACC_LEN=${DSV4_GOLDEN_AL} SGLANG_SIMULATE_ACC_METHOD=match-expected SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token"
             else
-                echo "WARNING: agentic MTP run (model=${MODEL_NAME}, DECODE_MTP_SIZE=${DECODE_MTP_SIZE}) has no golden AL wired in server_sglang.sh -- falling back to real (unsimulated, non-representative) acceptance. Add a case in server_sglang.sh and golden_al_distribution/ before shipping this arm. See golden_al_distribution/README.md." >&2
+                echo "WARNING: agentic spec-decoding run (model=${MODEL_NAME}, algorithm=${SPEC_DECODING:-mtp}, DECODE_MTP_SIZE=${DECODE_MTP_SIZE}) has no golden AL wired in server_sglang.sh -- falling back to real (unsimulated, non-representative) acceptance. Add a case in server_sglang.sh and golden_al_distribution/ before shipping this arm. See golden_al_distribution/README.md." >&2
             fi
         fi
     fi
@@ -1425,8 +1455,8 @@ else
 
         set +x
         decode_pid=$!
-        decode_pgid=$(ps -o pgid= -p "$decode_pid" 2>/dev/null | tr -d ' ')
-        : "${decode_pgid:=$decode_pid}"
+        decode_pgid=$decode_pid
+        SGLANG_OWNED_PGIDS+=("$decode_pgid")
     fi
 
     echo "Waiting for proxy server to be up..."
@@ -1455,8 +1485,7 @@ else
 
     echo "Killing the rank $RANK decode server"
     if [[ "$DRY_RUN" -eq 0 ]]; then
-        # Group-kill so TP-scheduler children release the tee pipe and the container exits.
-        kill -TERM -"${decode_pgid:-$decode_pid}" 2>/dev/null || true
+        exit 0
     fi
 
 fi
diff --git a/benchmarks/multi_node/amd_utils/stage_node_logs.sh b/benchmarks/multi_node/amd_utils/stage_node_logs.sh
new file mode 100755
index 0000000000..f1e67814e4
--- /dev/null
+++ b/benchmarks/multi_node/amd_utils/stage_node_logs.sh
@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+
+set -eo pipefail
+
+if [[ $# -ne 2 ]]; then
+    echo "Usage: $0 <node-local-log-dir> <shared-log-dir>" >&2
+    exit 2
+fi
+
+SOURCE_LOGS=$1
+SHARED_LOGS=$2
+
+if [[ ! -d "$SOURCE_LOGS" ]]; then
+    echo "[logs][ERROR] no node-local logs found on $(hostname): $SOURCE_LOGS" >&2
+    exit 1
+fi
+
+# Server containers create the source tree as root, and node 0 may have already
+# created the shared destination as root. The Slurm nodes provide passwordless
+# sudo for the same Docker lifecycle used by job.slurm.
+sudo mkdir -p "$SHARED_LOGS"
+sudo cp -r "$SOURCE_LOGS"/. "$SHARED_LOGS"/
+echo "[logs] staged $(hostname):$SOURCE_LOGS -> $SHARED_LOGS"
diff --git a/benchmarks/runtime_settings.sh b/benchmarks/runtime_settings.sh
index 0b04efe9ba..a369898f8b 100644
--- a/benchmarks/runtime_settings.sh
+++ b/benchmarks/runtime_settings.sh
@@ -32,4 +32,4 @@ export SGLANG_TORCH_PROFILER_DIR='/workspace'
 export VLLM_TORCH_PROFILER_DIR='/workspace'
 
 # Explicitly forward these settings across container boundaries.
-export INFERENCEX_RUNTIME_ENV_VARS="OPENAI_API_KEY SWEBENCH_EXPECTED_INSTANCES SWEBENCH_AGENT_STEP_LIMIT SWEBENCH_AGENT_TIMEOUT SWEBENCH_AGENT_EXIT_GRACE SWEBENCH_WATCHDOG_POLL SWEBENCH_SANDBOX_SWEEP SWEBENCH_SKIP_SCORE SWEBENCH_EVAL_TIMEOUT SWEBENCH_SCORE_TIMEOUT SWEBENCH_MAX_WORKERS EVAL_ENDPOINT_READY_TIMEOUT_SECONDS EVAL_MODEL_STABILIZATION_SECONDS AIPERF_FAILED_REQUEST_THRESHOLD AIPERF_LIVE_FAILED_REQUEST_THRESHOLD AIPERF_TRACE_IDLE_GAP_CAP_SECONDS AIPERF_PYTHON_VERSION AIPERF_WARMUP_REQUESTS_PER_LANE AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES AGENTIC_WARMUP_GRACE_PERIOD AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS AIPERF_UNSAFE_OVERRIDE ENABLE_AGENTX_POWER VLLM_ENGINE_READY_TIMEOUT_S SGLANG_TORCH_PROFILER_DIR VLLM_TORCH_PROFILER_DIR"
+export INFERENCEX_RUNTIME_ENV_VARS="OPENAI_API_KEY SWEBENCH_EXPECTED_INSTANCES SWEBENCH_AGENT_STEP_LIMIT SWEBENCH_AGENT_TIMEOUT SWEBENCH_AGENT_EXIT_GRACE SWEBENCH_WATCHDOG_POLL SWEBENCH_SANDBOX_SWEEP SWEBENCH_SKIP_SCORE SWEBENCH_EVAL_TIMEOUT SWEBENCH_SCORE_TIMEOUT SWEBENCH_MAX_WORKERS EVAL_ENDPOINT_READY_TIMEOUT_SECONDS EVAL_MODEL_STABILIZATION_SECONDS AIPERF_FAILED_REQUEST_THRESHOLD AIPERF_LIVE_FAILED_REQUEST_THRESHOLD AIPERF_TRACE_IDLE_GAP_CAP_SECONDS AIPERF_PYTHON_VERSION AIPERF_WARMUP_REQUESTS_PER_LANE AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES AGENTIC_WARMUP_GRACE_PERIOD AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS AIPERF_EXPERIMENTAL_FAST AIPERF_UNSAFE_OVERRIDE ENABLE_AGENTX_POWER REQUIRE_POWER VLLM_ENGINE_READY_TIMEOUT_S SGLANG_TORCH_PROFILER_DIR VLLM_TORCH_PROFILER_DIR"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index e233f73564..f291bce549 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1193,8 +1193,15 @@ minimaxm3-fp8-mi325x-vllm-agentic-mtp:
       search-space:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 10, 12, 14, 16, 18] }
 
-dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp:
-  image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911
+dsv4-fp4-mi355x-sglang-disagg-agentic-umbp-dspark:
+  # Renamed from dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp when this arm
+  # moved from EAGLE/MTP to DSpark; earlier perf-changelog entries are recorded
+  # under the old key.
+  # 20260913 rather than 20260911: it is the first tag carrying the DSpark
+  # optimizations, which is the whole point of this arm. CLIENT_IMAGE stays at
+  # 20260907 -- that container only runs the load generator, so the
+  # server-side DSpark work does not reach it.
+  image: lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260913
   model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:mi355x-amds
@@ -1207,7 +1214,7 @@ dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - spec-decoding: "mtp"
+      - spec-decoding: "draft_model"
         conc-list: [ 4 ]
         kv-offloading: none
         prefill:
@@ -1226,7 +1233,7 @@ dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp:
           additional-settings:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=3"
-      - spec-decoding: "mtp"
+      - spec-decoding: "draft_model"
         conc-list: [ 16 ]
         kv-offloading: none
         prefill:
@@ -1245,7 +1252,7 @@ dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp:
           additional-settings:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=3"
-      - spec-decoding: "mtp"
+      - spec-decoding: "draft_model"
         conc-list: [ 32, 48 ]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
@@ -1266,7 +1273,7 @@ dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp:
           additional-settings:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=3"
-      - spec-decoding: "mtp"
+      - spec-decoding: "draft_model"
         conc-list: [ 128, 192, 256 ]
         kv-offloading: dram
         kv-offload-backend: { name: umbp-linker }
diff --git a/docs/architecture.md b/docs/architecture.md
index a462f7215b..9e4b5c743a 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -277,7 +277,7 @@ The collector and reusable-artifact validator share format recognition, concurre
 
 Agentic throughput jobs have a different contract. They validate AIPerf output with [`infx/results/agentic/validate_agentic_result.py`](../infx/results/agentic/validate_agentic_result.py), upload an aggregate `bmk_agentic_<suffix>` artifact, and upload the raw `agentic_<suffix>` sibling containing trace-replay material. InferenceX-app pairs those siblings by their shared suffix. Agentic eval-only jobs follow the eval output contract instead and do not require a throughput result.
 
-Server logs and GPU metrics are diagnostic side artifacts. They are uploaded with `always()` so a failed run can still be investigated. Their presence does not turn a failed benchmark into a valid result.
+Server logs and GPU metrics are diagnostic side artifacts. They are uploaded with `always()` so a failed run can still be investigated. Their presence does not turn a failed benchmark into a valid result. On the AMD Slurm fleet, `/run_logs` is node-local; after the server step finishes, `job.slurm` merges the closed log tree from every allocated node into shared storage so the diagnostic artifact includes prefill and decode logs from the full deployment.
 
 ## Stage 6: artifact collection and handoff
 
diff --git a/docs/architecture_zh.md b/docs/architecture_zh.md
index 3e8fde840f..377a62cba4 100644
--- a/docs/architecture_zh.md
+++ b/docs/architecture_zh.md
@@ -277,7 +277,7 @@ rows = build_rows(raw_eval, metadata, source="eval_job/results.json")
 
 智能体吞吐量作业采用不同的契约。它们使用 [`infx/results/agentic/validate_agentic_result.py`](../infx/results/agentic/validate_agentic_result.py) 验证 AIPerf 输出，上传聚合的 `bmk_agentic_<suffix>` 工件，并上传包含追踪重放材料的原始 `agentic_<suffix>` 同级工件。InferenceX-app 通过它们共享的后缀对这些同级工件进行配对。智能体仅评测作业改为遵循评测输出契约，不要求吞吐量结果。
 
-服务器日志和 GPU 指标是诊断辅助工件。它们通过 `always()` 上传，因此失败的运行仍可供调查。它们的存在不会将失败的基准测试转变为有效结果。
+服务器日志和 GPU 指标是诊断辅助工件。它们通过 `always()` 上传，因此失败的运行仍可供调查。它们的存在不会将失败的基准测试转变为有效结果。在 AMD Slurm 机群上，`/run_logs` 是节点本地目录；服务器步骤结束后，`job.slurm` 会把每个已分配节点上已经关闭的日志树合并到共享存储中，使诊断工件包含整个部署的 Prefill 和 Decode 日志。
 
 ## 阶段 6：工件收集与交接
 
diff --git a/docs/recovery-results-procedures.md b/docs/recovery-results-procedures.md
index 0d140d4c84..8b8ac153ac 100644
--- a/docs/recovery-results-procedures.md
+++ b/docs/recovery-results-procedures.md
@@ -425,3 +425,28 @@ Remaining durable fix:
 ```
 
 This evidence is the completion gate. “Workflow green” without artifact identity, source/merge identity, and ingest counts is not a verified result recovery.
+
+### AMD multi-node SGLang teardown
+
+On exit, including a failed startup/readiness check, the AMD SGLang launcher sends
+TERM only to its recorded `setsid` process groups. Normal completion stages results
+before this cleanup. It allows 30 seconds for graceful
+exit, then sends KILL to surviving groups and checks for exit for another five
+seconds. This handles orphaned or TERM-resistant workers that otherwise hold log
+pipes open. These cleanup deadlines do not change profiling, evaluation, or server
+readiness deadlines. A failed client retains its exit status; unresolved cleanup
+fails an otherwise successful node. Kernel-blocked processes may still require
+separately authorized node repair. Do not change or discard completed metrics to
+work around teardown failures. A single EXIT handler owns group cleanup and the
+existing UMBP standalone PID cleanup; the latter still runs if group cleanup fails.
+
+### AMD multi-node GPU preflight coordination
+
+The Slurm launcher completes Docker pre-clean and the existing GPU VRAM drain
+check on every selected node in a separate Slurm step before launching any server
+container. A failed preflight prevents the serving step; it does not consume a
+healthy peer's container-readiness deadline. Node-local `preflight_<hostname>.log`
+files are included in the normal log fan-in, including failures. The VRAM threshold,
+15-minute GPU guard, and container/server readiness deadlines remain unchanged.
+This coordination prevents a peer-barrier race; it does not repair a GPU driver
+that fails to reclaim memory. The existing Docker pre-clean scope is unchanged.
diff --git a/docs/recovery-results-procedures_zh.md b/docs/recovery-results-procedures_zh.md
index df3ed5e045..1427077bb6 100644
--- a/docs/recovery-results-procedures_zh.md
+++ b/docs/recovery-results-procedures_zh.md
@@ -425,3 +425,23 @@ Remaining durable fix:
 ```
 
 这些证据就是完成关卡。如果没有制品身份、source/merge 身份和摄取数量，仅仅“工作流绿色”并不代表结果恢复已经验证。
+
+### AMD 多节点 SGLang 清理
+
+退出时（包括启动或就绪检查失败），AMD SGLang 启动器仅向其记录的 `setsid`
+进程组发送 TERM，等待最多 30 秒。正常完成时，先暂存结果再进行清理。随后向仍存活的进程组发送 KILL，再等待最多
+5 秒并检查退出状态。这可以清理已成为孤儿进程或忽略 TERM 的工作进程，避免其
+持续占用日志管道。这些清理期限不会改变性能采集、评估或服务器就绪检查的期限。
+客户端失败时保留原退出码；若客户端成功但清理仍未完成，则节点任务失败。
+内核阻塞的进程仍可能需要另行授权的节点修复。不要为绕过清理失败而修改或丢弃
+已完成的指标。单一 EXIT 处理器统一负责进程组清理和现有 UMBP 独立进程 PID
+清理；即使进程组清理失败，后者仍会执行。
+
+### AMD 多节点 GPU 预检协调
+
+Slurm 启动器先在独立步骤中完成所有选定节点的 Docker 预清理和现有 GPU VRAM
+回收检查，然后才启动服务器容器。任一节点预检失败都会阻止服务步骤启动，不会
+消耗健康节点等待容器就绪的期限。节点本地的 `preflight_<hostname>.log` 文件
+通过常规日志汇总流程收集，包括失败日志。VRAM 阈值、15 分钟 GPU 检查期限及
+容器和服务器就绪期限均保持不变。这一协调消除了节点间等待的竞态，但无法修复
+不能回收显存的 GPU 驱动。现有 Docker 预清理范围保持不变。
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f987183523..487f46dc69 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8003,3 +8003,18 @@
     - "Adds glm5.2/fp8 routing to runners/launch_b200-nscale-compat.sh (MODEL_PATH default /scratch/models/GLM-5.2-FP8, SRT_SLURM_MODEL_PREFIX glm5.2-fp8), which previously hard-failed with 'Unsupported model prefix/precision' for this model. The script's completeness-checked, flock-serialized hf download stages the checkpoint on the day-zero run if it is not already on the cluster."
     - "Re-pin from lmsysorg/sglang:nightly-dev-cu13-20260907-30705c00 to lmsysorg/sglang:nightly-dev-cu13-20260908-20ca564b (2026-09-08 cu13 dev nightly, digest sha256:9a352a35c973a2357372e85f3bcb5388b6b3c46c1329165987260f3b089647dc; Docker Hub last pushed 2026-09-08T01:40:59Z, tag commit sgl-project/sglang@20ca564b). The 2026-09-07 build carries an unguarded kv_index_translator.translate_dcp_read_ids call on the DSA fp8 KV read path that the EAGLE draft backend never binds, so GLM-5.2 MTP runs crash intermittently with AttributeError (observed on the MI355X FP8 sibling in run 34173459478 after 74 minutes of serving). sgl-project/sglang#38318 (merged 2026-09-07T20:03Z) adds the None guard and is six commits behind 20ca564b."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2863
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-disagg-agentic-umbp-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Switch all four arms from EAGLE (spec-decoding: mtp) to DSpark (spec-decoding: draft_model) at DECODE_MTP_SIZE=3, and bump the image from lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911 to -20260913, the first tag carrying the DSpark optimizations. CLIENT_IMAGE stays at 20260907 (load generator only). Topology, concurrency list, HiCache/UMBP settings, and precision are unchanged."
+    - "DeepSeek-V4-Pro-AgentX in models.yaml gains dspark_flags (--speculative-algorithm DSPARK --speculative-eagle-topk 1) alongside mtp_flags rather than replacing it, so recipes remaining on spec-decoding: mtp keep their EAGLE arm unchanged; dp_flags gains --enable-dp-lm-head, required by SGLang for DSpark under DP attention and inert for EAGLE. The -0813 checkpoint bundles the draft head, so no separate draft checkpoint is staged."
+    - "server_sglang.sh builds DSpark flags as --speculative-dspark-block-size DECODE_MTP_SIZE --speculative-num-steps 1 --speculative-num-draft-tokens (DECODE_MTP_SIZE + 1). DECODE_MTP_SIZE is the draft length for both algorithms but EAGLE spends it as sequential draft passes while DSpark emits one block, so num-steps is pinned to 1. The verify window is identical for both, so the MORI decode dispatch scaling by (DECODE_MTP_SIZE + 1) is unchanged. A model set to draft_model without dspark_flags now fails hard instead of silently serving EAGLE."
+    - "Golden AL is unchanged: the AgentX curve is keyed on checkpoint, thinking mode, and draft length, not on the algorithm, so DeepSeek-V4-Pro-0813 at draft length 3 keeps AL 3.01 from golden_al_distribution/dsv4-pro-0813-dspark.yaml (thinking_on)."
+    - "将四条臂全部从 EAGLE（spec-decoding: mtp）切换为 DSpark（spec-decoding: draft_model），DECODE_MTP_SIZE=3；镜像由 lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260911 升级到 -20260913（首个带 DSpark 优化的标签），CLIENT_IMAGE 保持 20260907（仅负载生成器）。拓扑、并发列表、HiCache/UMBP 设置与精度均不变。models.yaml 中 DeepSeek-V4-Pro-AgentX 新增 dspark_flags（与 mtp_flags 并列而非替换，因此仍用 mtp 的配方 EAGLE 分支不受影响），dp_flags 增加 --enable-dp-lm-head。server_sglang.sh 按 SPEC_DECODING == draft_model 分流：DSpark 一次前向出一整块，故 num-steps 钉死为 1，草稿长度转为 --speculative-dspark-block-size；验证窗口（长度+1）与 EAGLE 相同，所以 MORI 解码 dispatch 的 ×(DECODE_MTP_SIZE+1) 缩放无需改动。黄金 AL 不变：按检查点而非算法选表，DeepSeek-V4-Pro-0813 在草稿长度 3 仍为 AL 3.01。"
+    - "Rename the config key from dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp to dsv4-fp4-mi355x-sglang-disagg-agentic-umbp-dspark to reflect the DSpark algorithm and the UMBP-linker arm; earlier entries for this recipe are recorded under the old key. No other field changes with the rename."
+    - "配置键由 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp 改名为 dsv4-fp4-mi355x-sglang-disagg-agentic-umbp-dspark，以反映 DSpark 算法与 UMBP-linker 分支；该配方此前的条目仍记录在旧键名下。改名不伴随任何其他字段变化。"
+    - "Collect completed logs from every Slurm node so both prefill and decode artifacts are uploaded; forward workflow-owned AgentX fast-mode and power inputs, and supply the existing router port to evaluation."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3188