SemiAnalysisAI · CrimsonDump · Aug 13, 2026
diff --git a/benchmarks/multi_node/agentic/glm5.1_fp8_b200_tilert-disagg.sh b/benchmarks/multi_node/agentic/glm5.1_fp8_b200_tilert-disagg.sh
@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    DURATION \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    FRAMEWORK
+
+require_agentic_kv_offload_none
+
+export MODEL_NAME=glm5
+export TILERT_MODEL_TYPE=glm-5
+
+export DECODE_KV_DTYPE=fp8
+export PREFILL_KV_DTYPE=fp8_ds_mla
+
+export TILERT_PARSER=none
+
+exec bash "$(dirname "$0")/../tilert_utils/submit.sh"
diff --git a/benchmarks/multi_node/tilert_utils/run_node.sh b/benchmarks/multi_node/tilert_utils/run_node.sh
@@ -23,6 +23,18 @@ PREFILL_KV_DTYPE=${PREFILL_KV_DTYPE:-fp8_ds_mla}
 PREFILL_SPEC=(--speculative-config '{"method":"mtp","num_speculative_tokens":1}')
 DECODE_MTP=(--with-mtp)
 
+TILERT_IS_AGENTIC=0
+if [[ "${IS_AGENTIC:-0}" == "1" || "${SCENARIO_TYPE:-}" == "agentic-coding" ]]; then
+    TILERT_IS_AGENTIC=1
+fi
+
+if [[ "$TILERT_IS_AGENTIC" == "1" ]]; then
+    TILERT_QUEUE_TIMEOUT=${TILERT_QUEUE_TIMEOUT:-1800}
+fi
+TILERT_QUEUE_TIMEOUT=${TILERT_QUEUE_TIMEOUT:-0}
+
+AGENTIC_LOGS_DIR=${AGENTIC_LOGS_DIR:-$RESULT_DIR/LOGS/agentic}
+
 : "${DECODE_HOST:?DECODE_HOST is unset -- submit.sh must export it}"
 : "${PREFILL_HOST:?PREFILL_HOST is unset -- submit.sh must export it}"
 : "${TILERT_ROLE:?TILERT_ROLE is unset -- submit.sh must set it to decode or prefill}"
@@ -156,8 +168,10 @@ start_decode() {
 }
 
 start_prefill() {
+    local served=("$MODEL_NAME")
+    [[ -n "${MODEL:-}" && "$MODEL" != "$MODEL_NAME" ]] && served+=("$MODEL")
     local cmd=(vllm serve "$MODEL_PATH"
-        --served-model-name "$MODEL_NAME" --port "$PREFILL_PORT"
+        --served-model-name "${served[@]}" --port "$PREFILL_PORT"
         --tensor-parallel-size "$PREFILL_TP" --max-model-len "$MAX_MODEL_LEN"
         --enforce-eager --trust-remote-code --return-tokens-as-token-ids
         --gpu-memory-utilization "$GPU_MEM_UTIL" --kv-cache-dtype "$PREFILL_KV_DTYPE"
@@ -171,7 +185,8 @@ start_router() {
     local cmd=(env CUDA_VISIBLE_DEVICES= "${PY:-python}" -m tilert.pd_vllm.pd_router
         --vllm-url "http://$PREFILL_HOST:$PREFILL_PORT"
         --decode "$DECODE_HOST:$DECODE_CTRL_PORT:$DECODE_HTTP_PORT"
-        --port "$ROUTER_PORT" --model-path "$MODEL_PATH" --parser "$TILERT_PARSER")
+        --port "$ROUTER_PORT" --model-path "$MODEL_PATH" --parser "$TILERT_PARSER"
+        --queue-timeout "$TILERT_QUEUE_TIMEOUT")
     log_and_run_bg router "$BENCHMARK_LOGS_DIR/tilert_router.log" "${cmd[@]}"
     ROUTER_PID=$LAST_BG_PID
 }
@@ -211,16 +226,37 @@ run_bench_and_eval() {
             --result-filename "$(bench_result_stem "$conc")" --result-dir "$RESULT_DIR" \
             || { rc=$?; echo "[bench] WARNING: conc=$conc failed/timed out (rc=$rc)"; }
     done
-    if [[ "${RUN_EVAL}" = "true" ]]; then
-        if [[ -n "${EVAL_CONC:-}" ]]; then
-            export EVAL_CONCURRENT_REQUESTS="$EVAL_CONC"
-        else
-            export EVAL_CONCURRENT_REQUESTS="$(tr ' ' '\n' <<< "$CONC_LIST" | sort -n | tail -1)"
-        fi
-        export CONC="$EVAL_CONCURRENT_REQUESTS"
-        run_eval --framework lm-eval --port "$ROUTER_PORT"
-        append_lm_eval_summary
+    run_lm_eval
+    return $rc
+}
+
+run_lm_eval() {
+    [[ "${RUN_EVAL}" = "true" ]] || return 0
+    if [[ -n "${EVAL_CONC:-}" ]]; then
+        export EVAL_CONCURRENT_REQUESTS="$EVAL_CONC"
+    else
+        export EVAL_CONCURRENT_REQUESTS="$(tr ' ' '\n' <<< "$CONC_LIST" | sort -n | tail -1)"
     fi
+    export CONC="$EVAL_CONCURRENT_REQUESTS"
+    run_eval --framework lm-eval --port "$ROUTER_PORT"
+    append_lm_eval_summary
+}
+
+run_agentic_replay() {
+    wait_for_server_ready --port "$ROUTER_PORT" \
+        --server-log "$BENCHMARK_LOGS_DIR/tilert_router.log" --server-pid "$ROUTER_PID"
+    local rc=0 conc conc_result_dir
+    local result_filename_base="$RESULT_FILENAME"
+    for conc in $CONC_LIST; do
+        conc_result_dir="$AGENTIC_LOGS_DIR/conc_${conc}"
+        mkdir -p "$conc_result_dir"
+        export CONC="$conc"
+        export RESULT_FILENAME="${result_filename_base}_conc${conc}"
+        build_replay_cmd "$conc_result_dir"
+        run_agentic_replay_and_write_outputs "$conc_result_dir" \
+            || { rc=$?; echo "[agentic] WARNING: conc=$conc failed/timed out (rc=$rc)"; }
+    done
+    export RESULT_FILENAME="$result_filename_base"
     return $rc
 }
 
@@ -247,13 +283,21 @@ case "$TILERT_ROLE" in
     prefill)
         rdma_preflight || exit 1
         rm -f "$DONE_SENTINEL"
+        if [[ "$TILERT_IS_AGENTIC" == "1" ]]; then
+            resolve_trace_source
+            install_agentic_deps
+        fi
         wait_for_tcp "$DECODE_HOST" "$DECODE_CTRL_PORT" "$DECODE_WAIT" \
             || echo "[prefill] WARNING: timed out waiting for the decode ctrl port ($DECODE_HOST:$DECODE_CTRL_PORT), starting anyway"
         start_prefill
         wait_for_tcp "$PREFILL_HOST" "$PREFILL_PORT" "${PREFILL_WAIT:-3600}" \
             || echo "[prefill] WARNING: timed out waiting for the vLLM port ($PREFILL_HOST:$PREFILL_PORT), continuing (see $BENCHMARK_LOGS_DIR/tilert_prefill.log)"
         start_router
-        run_bench_and_eval; BENCH_RC=$?
+        if [[ "$TILERT_IS_AGENTIC" == "1" ]]; then
+            run_agentic_replay; BENCH_RC=$?
+        else
+            run_bench_and_eval; BENCH_RC=$?
+        fi
         touch "$DONE_SENTINEL"
         kill "$ROUTER_PID" "$PREFILL_PID" 2>/dev/null || true
         exit $BENCH_RC

diff --git a/benchmarks/multi_node/tilert_utils/setup_deps.sh b/benchmarks/multi_node/tilert_utils/setup_deps.sh
@@ -1,6 +1,6 @@
 #!/bin/bash
 
-TILERT_VERSION="${TILERT_VERSION:-0.1.5.post2}"
+TILERT_VERSION="${TILERT_VERSION:-0.1.5.post3}"
 TILERT_PIP_INDEX_URL="${TILERT_PIP_INDEX_URL:-}"
 
 TILERT_HTTP_DEPS="${TILERT_HTTP_DEPS:-fastapi uvicorn httpx}"

@@ -8834,7 +8834,7 @@ glm5.1-fp8-b200-tilert:
   runner: b200-multinode
   precision: fp8
   framework: tilert
-  router: { name: tilert-pd-router, version: "0.1.5.post2" }
+  router: { name: tilert-pd-router, version: "0.1.5.post3" }
   multinode: true
   disagg: true
   kv-p2p-transfer: nixl
@@ -8880,3 +8880,35 @@ glm5.1-fp8-b200-tilert:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
+
+glm5.1-fp8-b200-tilert-agentic:
+  image: ghcr.io/tile-ai/tilert:0.1.5
+  model: zai-org/GLM-5.1-FP8
+  model-prefix: glm5.1
+  runner: cluster:b200-dgxc-slurm
+  precision: fp8
+  framework: tilert
+  router: { name: tilert-pd-router, version: "0.1.5.post3" }
+  multinode: true
+  disagg: true
+  kv-p2p-transfer: nixl
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: "mtp"
+        conc-list: [1]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_IMAGE=vllm/vllm-openai:v0.26.0"
+          - "PREFILL_NODES=1"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
diff --git a/configs/runners.yaml b/configs/runners.yaml
@@ -80,6 +80,10 @@ labels:
     - b200-dgxc-slurm_7
     - b200-dgxc-slurm_8
     - b200-dgxc-slurm_9
+  cluster:b200-dgxc-slurm:
+    - b200-dgxc-slurm_7
+    - b200-dgxc-slurm_8
+    - b200-dgxc-slurm_9
   b200-new:
     - b200-nscale-slurm_0
     - b200-nscale-slurm_1

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
@@ -6076,3 +6076,14 @@
     - "Add GB200 DeepSeek-V4-Pro FP4 Dynamo-vLLM AgentX mirroring the GB300 PR #2571 MTP tuning, with every GB300 4-GPU worker sized to 8 GPUs on GB200."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2636
 
+- config-keys:
+    - glm5.1-fp8-b200-tilert-agentic
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add the agentic-coding scenario for GLM-5.1 FP8 B200 TileRT PD-disaggregation, reusing the fixed-seq-len topology (1 prefill node TP8 + 1 decode node TP8, NIXL KV transfer, MTP) with an AIPerf trace replay in place of benchmark_serving.py"
+    - "Separate config key rather than a second scenario on glm5.1-fp8-b200-tilert: agentic master configs must declare an exact cluster:<name> runner, so this entry adds cluster:b200-dgxc-slurm to configs/runners.yaml over the same three b200-dgxc-slurm nodes the b200-multinode label already uses"
+    - "TileRT decode serves one sequence at a time, so conc-list is the single point [1]; the router waits for a free decode node (--queue-timeout) instead of answering 429 when a session fans out into concurrent sub-conversations"
+    - "kv-offloading: none -- TileRT keeps all KV state on the GPU and exposes no DRAM offload tier"
+    - "GLM-5.1 has 202752 native positions, so AIPerf's --max-context-length drops the corpus traces whose peak context exceeds it: 175 of the 393 traces in cc-traces-weka-062126-256k remain eligible"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2645
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
@@ -138,7 +138,9 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         export UCX_NET_DEVICES="${UCX_NET_DEVICES:-mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1}"
         export UCX_MEMTYPE_CACHE="${UCX_MEMTYPE_CACHE:-n}"
         export UCX_MEMTYPE_REG_WHOLE="${UCX_MEMTYPE_REG_WHOLE:-n}"
-        TILERT_DISAGG="$GITHUB_WORKSPACE/benchmarks/multi_node/${EXP_NAME%%_*}_${PRECISION}_b200_${FRAMEWORK}-disagg.sh"
+        TILERT_SUBDIR="multi_node"
+        [[ "${SCENARIO_SUBDIR}" == "agentic/" ]] && TILERT_SUBDIR="multi_node/agentic"
+        TILERT_DISAGG="$GITHUB_WORKSPACE/benchmarks/${TILERT_SUBDIR}/${EXP_NAME%%_*}_${PRECISION}_b200_${FRAMEWORK}-disagg.sh"
         [[ -f "$TILERT_DISAGG" ]] || { echo "tilert disagg script not found: $TILERT_DISAGG"; exit 1; }
         exec bash "$TILERT_DISAGG"
         exit 1