From da35b40f831baaff835882c1c743a64465364789 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 15:26:51 -0500
Subject: [PATCH 1/7] refactor(tilert): port MI355X AgentX to native srt-slurm

---
 .../agentic/glm5.3_fp8_mi355x_tilert.sh       | 172 ------
 .../benchmarks/multi_node/amd_utils/env.sh    |  27 +-
 .../benchmarks/multi_node/amd_utils/job.slurm |  55 +-
 .../multi_node/amd_utils/models_tilert.yaml   |   6 -
 .../benchmarks/multi_node/amd_utils/server.sh |   3 -
 .../multi_node/amd_utils/server_atom.sh       |   1 -
 .../multi_node/amd_utils/server_sglang.sh     |   1 -
 .../multi_node/amd_utils/server_tilert.sh     | 573 ------------------
 .../multi_node/amd_utils/setup_deps.sh        | 136 -----
 .../configs/glm5.3-tilert-rocm.sh             |  68 +++
 .../agentx/disagg-1p1d-tp8-mtp.yaml           | 103 ++++
 inferencex-e2e/configs/amd-master.yaml        |  23 +-
 inferencex-e2e/perf-changelog.yaml            |   6 +
 inferencex-e2e/runners/launch_mi355x-amds.sh  |   6 +-
 14 files changed, 192 insertions(+), 988 deletions(-)
 delete mode 100644 inferencex-e2e/benchmarks/multi_node/agentic/glm5.3_fp8_mi355x_tilert.sh
 delete mode 100644 inferencex-e2e/benchmarks/multi_node/amd_utils/models_tilert.yaml
 delete mode 100644 inferencex-e2e/benchmarks/multi_node/amd_utils/server_tilert.sh
 delete mode 100644 inferencex-e2e/benchmarks/multi_node/amd_utils/setup_deps.sh
 create mode 100644 inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
 create mode 100644 inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml

diff --git a/inferencex-e2e/benchmarks/multi_node/agentic/glm5.3_fp8_mi355x_tilert.sh b/inferencex-e2e/benchmarks/multi_node/agentic/glm5.3_fp8_mi355x_tilert.sh
deleted file mode 100644
index 82e81aee32..0000000000
--- a/inferencex-e2e/benchmarks/multi_node/agentic/glm5.3_fp8_mi355x_tilert.sh
+++ /dev/null
@@ -1,172 +0,0 @@
-#!/usr/bin/env bash
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-source "$SCRIPT_DIR/../../benchmark_lib.sh"
-
-check_env_vars \
-    CONC_LIST \
-    ISL \
-    OSL \
-    IMAGE \
-    SPEC_DECODING \
-    MODEL_PATH \
-    MODEL_NAME \
-    PREFILL_NUM_WORKERS \
-    PREFILL_TP \
-    PREFILL_EP \
-    PREFILL_DP_ATTN \
-    DECODE_NUM_WORKERS \
-    DECODE_TP \
-    DECODE_EP \
-    DECODE_DP_ATTN \
-    PREFILL_NODES \
-    DECODE_NODES \
-    RANDOM_RANGE_RATIO \
-    DURATION \
-    MODEL_PREFIX \
-    PRECISION \
-    RESULT_FILENAME \
-    KV_OFFLOADING \
-    IS_AGENTIC \
-    FRAMEWORK \
-    PREFILL_IMAGE
-
-if [[ -n "$SLURM_JOB_ID" ]]; then
-  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
-fi
-
-set -x
-
-cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
-
-export TIME_LIMIT=08:00:00
-export MODEL_PATH=$MODEL_PATH
-export MODEL_NAME=$MODEL_NAME
-export CONTAINER_IMAGE=$IMAGE
-export PREFILL_IMAGE
-
-export RESULT_FILENAME
-
-if [[ "$PREFILL_NODES" -ne 1 || "$DECODE_NODES" -ne 1 || \
-      "$PREFILL_NUM_WORKERS" -ne 1 || "$DECODE_NUM_WORKERS" -ne 1 ]]; then
-    echo "Error: tilert supports exactly 1 prefill node/worker + 1 decode node/worker" \
-         "(got PREFILL_NODES=$PREFILL_NODES x$PREFILL_NUM_WORKERS, DECODE_NODES=$DECODE_NODES x$DECODE_NUM_WORKERS)" >&2
-    exit 1
-fi
-
-if [[ "$KV_OFFLOADING" != "none" ]]; then
-    echo "Error: tilert has no KV offload backend; kv-offloading must be 'none' (got '$KV_OFFLOADING')" >&2
-    exit 1
-fi
-
-# TileRT configuration. Every value is explicit here: server_tilert.sh
-# validates each one with check_env_vars and supplies no defaults of its own.
-export TILERT_VERSION=0.1.6.post2
-export TILERT_PROFILE=glm5_2          # decode_server --model (TileRT model profile)
-export TILERT_MODEL_TYPE=glm-5        # weight_converter --model_type (fallback converter)
-export TILERT_MODEL_PKG=glm_5_2_rocm  # per-model converter package, preferred when importable
-export SERVED_MODEL_NAME=glm5_2
-# GLM-5.3's full context window (config.json max_position_embeddings), as every
-# in-tree GLM-5.2 recipe serves. (202752 was GLM-5.1's, inherited from the B200
-# TileRT recipe this mirrors.)
-#
-# Memory at this context, per rank, bf16 wire layout (verified against the
-# tilert 0.1.6 and vLLM 0.24.0 sources and the MI355X logs, 287.98 GiB cards;
-# the undivided PD buffer sizes are what 0.1.6 allocated on one card):
-#   decode : weights 90.72 GiB + engine cache window 93.25 GiB
-#            + PD receive buffer 99.06 GiB (receive_server.py, dense in max_seq_len)
-#   prefill: weights 90.45 GiB + profiling/non-torch 40.3 GiB + vLLM KV 91.71 GiB
-#            + PD staging buffer 99.06 GiB (prefill_connector.py, TP rank 0,
-#              allocated OUTSIDE vLLM's gpu-memory-utilization budget)
-# Undivided, neither side starts: the decode rank is node-marginal (~283 of
-# 288 GiB) and the prefill rank cannot fit at any utilization (~321 GiB).
-# tilert 0.1.6.post1 keeps both buffers on the GPU but shards them by layer
-# across the eight devices (layer lid on device lid % 8, TILERT_PD_SHARDS,
-# default on), so each card holds 12.54 GiB instead of 99.06 GiB on one.
-# convert() dequantises each layer on the device that received it, which spreads
-# its transients too (108.6 KiB/token, 82.9 GiB at 800k tokens) instead of
-# leaving them on cuda:0. Measured on 2x8 MI350X at this context with bf16 KV:
-# decode peaks at 202.1 GiB per card, prefill at 269.2 GiB of 287.69 GiB, and
-# the KV path stays device-to-device at 108 GB/s (81 GB in 751 ms, 54% of the
-# 4x400 GbE line rate). No host hop and no patch: the wheel runs as shipped.
-export TILERT_MAX_MODEL_LEN=1048576
-export TILERT_TRANSPORT=mooncake
-export TILERT_PARSER=none
-export TILERT_RDMA_STRICT=0
-export TILERT_CONVERT_LOCK_WAIT=21600
-export TILERT_SIMULATE_ACC_METHOD=match-expected
-export TILERT_WEIGHTS_DIR="/models/${MODEL_NAME}-tilert-tp${DECODE_TP}"
-# bf16 MLA KV on both roles. This is the only layout TileRT 0.1.6 can consume
-# from vLLM on ROCm: MlaNsaProfile.classify_layers infers the layout from the
-# cache tensor stride and accepts exactly 1152 B/token (bf16) or 656 B/token
-# (fp8_ds_mla). vLLM's ROCM_AITER_MLA_SPARSE backend has no fp8_ds_mla; its
-# plain "fp8" writes a flat 576 B/token row, which the connector rejects at
-# register_kv_caches. Explicit bfloat16 rather than auto so the stride does not
-# depend on the model dtype. Never float16: it passes the 1152 B check and is
-# then read as bf16.
-export PREFILL_KV_DTYPE=bfloat16
-# The ROCm backend supports block sizes [1, 64] and vLLM picks 1, which makes
-# the connector's KI plane copy fail and MLA address the wrong rows.
-export PREFILL_BLOCK_SIZE=64
-export DECODE_KV_DTYPE=bf16
-# The PD staging shard sits outside vLLM's budget, so vLLM needs 90.45 (weights)
-# + 40.3 (profiling) + 91.71 GiB (KV for one 1048576-token request) = 222.5 GiB
-# inside it: 0.85 x 287.98 = 244.8 GiB leaves 22 GiB of KV margin and 43 GiB
-# outside the budget for the 12.54 GiB staging shard plus the ~6.3 GiB non-torch
-# baseline measured on the decode OOM node (287.98 - 95.94 free - 184.17 - 1.58
-# reserved). 0.75 (216 GiB) refuses with "91.71 GiB KV cache is needed ...
-# available 85.25 GiB".
-export GPU_MEM_UTIL=0.85
-export SKIP_CONTAINER_BARRIER=0
-# Two images, one per rank, ~32 GB each. On a node that has neither cached the
-# pull alone outlasts the SGLang path's 300s default and the 1800s this script
-# used to hardcode, and the rank that comes up first waits out the whole
-# timeout while its peer is still pulling.
-export CONTAINER_BARRIER_TIMEOUT=5400
-export ROUTER_PORT=30000
-export PREFILL_PORT=8000
-export DECODE_CTRL_PORT=5556
-export DECODE_HTTP_PORT=5557
-export DECODE_WAIT=7200               # prefill waits for the decode ctrl port
-export PREFILL_WAIT=3600              # prefill waits for its own vLLM port
-export ROUTER_WAIT=10800              # decode waits for the router port to open
-
-if [[ "$SPEC_DECODING" == "mtp" ]]; then
-    # TileRT decode drafts at depth 3 (the only depth the ROCm GLM profile
-    # builds) and the golden acceptance curve is keyed on it. The vLLM prefill
-    # rank only has to materialise the MTP layer's KV, so it runs at 1.
-    export DECODE_MTP_SIZE=3
-    export PREFILL_SPEC_TOKENS=1
-else
-    export DECODE_MTP_SIZE=0
-    export PREFILL_SPEC_TOKENS=0
-fi
-export TILERT_QUEUE_TIMEOUT=1800  # requests wait on the bs=1 decode engine
-export THINKING_MODE=thinking_on
-
-if [[ "$PREFILL_EP" -ne 1 || "$DECODE_EP" -ne 1 || \
-      "$PREFILL_DP_ATTN" == "true" || "$DECODE_DP_ATTN" == "true" ]]; then
-    echo "Error: tilert runs pure TP8 on both roles; ep must be 1 and dp-attn false" >&2
-    exit 1
-fi
-export PREFILL_ENABLE_EP=false
-export PREFILL_ENABLE_DP=false
-export DECODE_ENABLE_EP=false
-export DECODE_ENABLE_DP=false
-
-JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
-    $PREFILL_NUM_WORKERS \
-    $DECODE_NODES \
-    $DECODE_NUM_WORKERS \
-    $ISL $OSL "${CONC_LIST// /x}" inf \
-    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
-    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
-    ${PREFILL_TP} ${DECODE_TP} \
-    ${RANDOM_RANGE_RATIO})
-
-if [[ $? -ne 0 ]]; then
-    echo "Failed to submit job" >&2
-    exit 1
-fi
-
-echo "$JOB_ID"
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/env.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/env.sh
index ba48879cb4..04adb875ac 100755
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/env.sh
+++ b/inferencex-e2e/benchmarks/multi_node/amd_utils/env.sh
@@ -2,21 +2,12 @@
 
 source "$(dirname "${BASH_SOURCE[0]}")/../../benchmark_lib.sh" --validation-only
 check_env_vars ENGINE
-# MoRI-IO queue-pair tuning, the UCX RoCE GID index, SGLang router logging and the
-# SGLang decode cuda-graph NCCL workaround. Only the SGLang MoRI KV path
-# below reads these. ENGINE=tilert moves KV over mooncake and starts no SGLang
-# router, so it is neither given nor reads them: validating them there would force
-# the recipe to invent MoRI tuning for a transport it never uses.
-if [[ "$ENGINE" != "tilert" ]]; then
-    check_env_vars \
-        MORI_IO_SQ_BACKOFF_TIMEOUT_US MORI_IO_QP_MAX_SEND_WR MORI_IO_QP_MAX_CQE MORI_IO_QP_MAX_SGE MORI_IO_TC_DISABLE \
-        UCX_IB_GID_INDEX MORI_APP_LOG_LEVEL SGLANG_ROUTER_STDOUT_LOGS TORCH_NCCL_BLOCKING_WAIT NCCL_BLOCKING_WAIT \
-        SGLANG_OPT_USE_AITER_INDEXER
-fi
-# Dual-engine environment setup for multi-node disaggregated serving.
-#
-# ENGINE=sglang-disagg or tilert selects the engine-specific block.
-#
+# SGLang MoRI environment.
+check_env_vars \
+    MORI_IO_SQ_BACKOFF_TIMEOUT_US MORI_IO_QP_MAX_SEND_WR MORI_IO_QP_MAX_CQE MORI_IO_QP_MAX_SGE MORI_IO_TC_DISABLE \
+    UCX_IB_GID_INDEX MORI_APP_LOG_LEVEL SGLANG_ROUTER_STDOUT_LOGS TORCH_NCCL_BLOCKING_WAIT NCCL_BLOCKING_WAIT \
+    SGLANG_OPT_USE_AITER_INDEXER
+
 # REQUIRED ENVIRONMENT VARIABLES:
 #   IBDEVICES - RDMA/InfiniBand device names (e.g., ionic_0,ionic_1,... or mlx5_0,mlx5_1,...)
 #               Set by runner or auto-detected from hostname.
@@ -119,10 +110,6 @@ else
     fi
 fi
 
-if [[ "$ENGINE" == "tilert" ]]; then
-    echo "[INFO] tilert: IBDEVICES=$IBDEVICES  NCCL_SOCKET_IFNAME=$NCCL_SOCKET_IFNAME  NCCL_IB_HCA=$NCCL_IB_HCA"
-
-else
 
     export SGLANG_USE_AITER=1
     export AITER_LOG_LEVEL=ERROR
@@ -237,5 +224,3 @@ else
             export GPU_MAX_HW_QUEUES=2
         fi
     fi
-
-fi
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/job.slurm b/inferencex-e2e/benchmarks/multi_node/amd_utils/job.slurm
index cecc989a0f..ef46cfc859 100755
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/job.slurm
+++ b/inferencex-e2e/benchmarks/multi_node/amd_utils/job.slurm
@@ -36,8 +36,6 @@ echo ""
 # at runtime, but the CWD remains the submit-time directory (amd_utils/).
 if [[ "$ENGINE" == "atom-disagg" ]]; then
     MODELS_YAML="$(pwd)/models_atom.yaml"
-elif [[ "$ENGINE" == "tilert" ]]; then
-    MODELS_YAML="$(pwd)/models_tilert.yaml"
 else
     MODELS_YAML="$(pwd)/models.yaml"
 fi
@@ -52,16 +50,6 @@ if [[ -z "${DOCKER_IMAGE_NAME:-}" ]]; then
     exit 1
 fi
 
-if [[ "$ENGINE" == "tilert" && -z "${PREFILL_IMAGE:-}" ]]; then
-    echo "Error: ENGINE=tilert requires PREFILL_IMAGE (e.g. PREFILL_IMAGE=vllm/vllm-openai-rocm:nightly-<sha> in prefill.additional-settings)."
-    exit 1
-fi
-if [[ "$ENGINE" == "tilert" ]]; then
-    # server_tilert.sh takes the container-creation barrier timeout from the
-    # recipe (no 300s default as on the SGLang path); fail here, before sbatch
-    # work is done, rather than inside the container.
-    check_env_vars CONTAINER_BARRIER_TIMEOUT
-fi
 
 # Resolve the models.yaml entry the same way server_sglang.sh does: agentic runs
 # (IS_AGENTIC) use the '<model>-AgentX' recipe, non-agentic disaggregated runs use
@@ -564,41 +552,6 @@ if [[ "$ENGINE" == "atom-disagg" ]]; then
         -e EXTRA_SERVER_ARGS=\${EXTRA_SERVER_ARGS:-}
         -e IBDEVICES=${IBDEVICES:-}
     )
-elif [[ "$ENGINE" == "tilert" ]]; then
-    DOCKER_ENV_ENGINE=(
-        -e MODEL_PATH=$DOCKER_MODEL_PATH
-        -e PREFILL_IMAGE=${PREFILL_IMAGE}
-        -e TILERT_VERSION=${TILERT_VERSION}
-        -e TILERT_PROFILE=${TILERT_PROFILE}
-        -e TILERT_MODEL_TYPE=${TILERT_MODEL_TYPE}
-        -e TILERT_MODEL_PKG=${TILERT_MODEL_PKG}
-        -e TILERT_MAX_MODEL_LEN=${TILERT_MAX_MODEL_LEN}
-        -e TILERT_TRANSPORT=${TILERT_TRANSPORT}
-        -e TILERT_PARSER=${TILERT_PARSER}
-        -e TILERT_QUEUE_TIMEOUT=${TILERT_QUEUE_TIMEOUT}
-        -e TILERT_WEIGHTS_DIR=${TILERT_WEIGHTS_DIR}
-        -e TILERT_RDMA_STRICT=${TILERT_RDMA_STRICT}
-        -e TILERT_CONVERT_LOCK_WAIT=${TILERT_CONVERT_LOCK_WAIT}
-        -e TILERT_SIMULATE_ACC_METHOD=${TILERT_SIMULATE_ACC_METHOD}
-        -e \"TILERT_EXTRA_ENV=${TILERT_EXTRA_ENV:-}\"
-        -e SERVED_MODEL_NAME=${SERVED_MODEL_NAME}
-        -e PREFILL_KV_DTYPE=${PREFILL_KV_DTYPE}
-        -e PREFILL_BLOCK_SIZE=${PREFILL_BLOCK_SIZE}
-        -e PREFILL_SPEC_TOKENS=${PREFILL_SPEC_TOKENS}
-        -e DECODE_KV_DTYPE=${DECODE_KV_DTYPE}
-        -e GPU_MEM_UTIL=${GPU_MEM_UTIL}
-        -e PREFILL_PORT=${PREFILL_PORT}
-        -e DECODE_CTRL_PORT=${DECODE_CTRL_PORT}
-        -e DECODE_HTTP_PORT=${DECODE_HTTP_PORT}
-        -e DECODE_WAIT=${DECODE_WAIT}
-        -e PREFILL_WAIT=${PREFILL_WAIT}
-        -e ROUTER_WAIT=${ROUTER_WAIT}
-        -e SKIP_CONTAINER_BARRIER=${SKIP_CONTAINER_BARRIER}
-        # Golden-acceptance selection on agentic MTP runs; unset elsewhere.
-        -e THINKING_MODE=${THINKING_MODE:-}
-        -e IBDEVICES=${IBDEVICES:-}
-        -e PYTHONPYCACHEPREFIX=/tmp/pycache
-    )
 else
     DOCKER_ENV_ENGINE=(
         -e SGLANG_WS_PATH=${WS_PATH}
@@ -751,12 +704,6 @@ else
 fi
 fi  # end: if ENGINE == atom-disagg
 
-RANK_IMAGE=
-if [[ \"$ENGINE\" == \"tilert\" && \"\$SLURM_PROCID\" -lt \"$xP\" ]]; then
-    RANK_IMAGE=\"$PREFILL_IMAGE\"
-    echo \"[tilert] rank \$SLURM_PROCID is a prefill rank; using PREFILL_IMAGE=\$RANK_IMAGE\"
-fi
-
 exec \$DOCKER_CMD run \
     --init \
     --stop-timeout 10 \
@@ -798,7 +745,7 @@ exec \$DOCKER_CMD run \
     ${CLIENT_DOCKER_ENV} \
     --name \"$DOCKER_CONT_NAME\" \
     --entrypoint \"\" \
-    \"\${RANK_IMAGE:-$DOCKER_IMAGE_NAME}\" bash -lc '
+    \"$DOCKER_IMAGE_NAME\" bash -lc '
         set -o pipefail
         mkdir -p /run_logs/slurm_job-'\"\$SLURM_JOB_ID\"'
         '"$RUN_FILE_FULL"' 2>&1 | tee /run_logs/slurm_job-'\"\$SLURM_JOB_ID\"'/server_\$(hostname).log
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/models_tilert.yaml b/inferencex-e2e/benchmarks/multi_node/amd_utils/models_tilert.yaml
deleted file mode 100644
index 1697059d54..0000000000
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/models_tilert.yaml
+++ /dev/null
@@ -1,6 +0,0 @@
-# Model-owned engine settings for ENGINE=tilert. Everything the caller owns
-# (profile, topology, ports, dtypes, draft depth) is set by the recipe, not here.
-GLM-5.3:
-  prefill_env: "VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 VLLM_ENGINE_READY_TIMEOUT_S=10800"
-  prefill_extra_flags: ""
-  decode_extra_flags: ""
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/server.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/server.sh
index f304f53b40..a693fb92cc 100755
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/server.sh
+++ b/inferencex-e2e/benchmarks/multi_node/amd_utils/server.sh
@@ -5,7 +5,6 @@ source "$(dirname "${BASH_SOURCE[0]}")/../../benchmark_lib.sh" --validation-only
 # Dispatches to the engine-specific server launcher based on ENGINE env var.
 #   ENGINE=sglang-disagg (default) -> server_sglang.sh (SGLang + MoRI)
 #   ENGINE=atom-disagg             -> server_atom.sh (ATOM + mooncake)
-#   ENGINE=tilert                  -> server_tilert.sh (vLLM prefill + TileRT decode)
 
 check_env_vars ENGINE WS_PATH
 if [[ -f /config/hicache_mc.env ]]; then
@@ -20,8 +19,6 @@ echo "[DISPATCHER] ENGINE=$ENGINE  WS_PATH=$WS_PATH"
 if [[ "$ENGINE" == "atom-disagg" ]]; then
     export ATOM_WS_PATH="$WS_PATH"
     source "$WS_PATH/server_atom.sh"
-elif [[ "$ENGINE" == "tilert" ]]; then
-    source "$WS_PATH/server_tilert.sh"
 else
     source "$WS_PATH/server_sglang.sh"
 fi
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_atom.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/server_atom.sh
index caa187b6bc..4cc51efea8 100755
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_atom.sh
+++ b/inferencex-e2e/benchmarks/multi_node/amd_utils/server_atom.sh
@@ -16,7 +16,6 @@ check_env_vars \
 
 EXTRA_SERVER_ARGS="${EXTRA_SERVER_ARGS:-}"
 
-source $ATOM_WS_PATH/setup_deps.sh
 source $ATOM_WS_PATH/env_atom.sh
 
 # lm-eval with high num_concurrent exhausts the default 1024 FD limit.
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_sglang.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/server_sglang.sh
index bd6403f8c9..571d1799e0 100755
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_sglang.sh
+++ b/inferencex-e2e/benchmarks/multi_node/amd_utils/server_sglang.sh
@@ -20,7 +20,6 @@ BENCH_MAX_CONC_VALUE=$(echo "$BENCH_MAX_CONCURRENCY" | tr 'x' '\n' | sort -n | t
 # can resolve formulas like "BENCH_MAX_CONC_VALUE*2" for max_running_requests.
 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
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_tilert.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/server_tilert.sh
deleted file mode 100644
index 3a66c926d7..0000000000
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/server_tilert.sh
+++ /dev/null
@@ -1,573 +0,0 @@
-#!/bin/bash
-# TileRT disaggregated launcher: upstream vLLM ROCm prefill (TileRTConnector,
-# kv_producer) + TileRT decode_server + the OpenAI-compatible pd_router.
-# Every value below is supplied by the recipe through job.slurm; this script
-# validates them and never invents a default for caller-owned configuration.
-
-source "$(dirname "${BASH_SOURCE[0]}")/../../benchmark_lib.sh" --validation-only
-
-check_env_vars \
-    NODE0_ADDR NODE_RANK MODEL_DIR MODEL_NAME MODEL_PATH xP yD IPADDRS \
-    DRY_RUN GPUS_PER_NODE PREFILL_TP_SIZE DECODE_TP_SIZE \
-    BENCH_INPUT_LEN BENCH_OUTPUT_LEN BENCH_MAX_CONCURRENCY \
-    RUN_EVAL EVAL_ONLY EVAL_FRAMEWORK BENCHMARK_LOGS_DIR WS_PATH \
-    SLURM_JOB_ID SPEC_DECODING \
-    TILERT_PROFILE TILERT_MODEL_TYPE TILERT_MODEL_PKG TILERT_MAX_MODEL_LEN \
-    TILERT_TRANSPORT TILERT_PARSER TILERT_QUEUE_TIMEOUT TILERT_WEIGHTS_DIR \
-    TILERT_RDMA_STRICT TILERT_CONVERT_LOCK_WAIT TILERT_SIMULATE_ACC_METHOD \
-    PREFILL_KV_DTYPE PREFILL_BLOCK_SIZE PREFILL_SPEC_TOKENS DECODE_KV_DTYPE \
-    DECODE_MTP_SIZE GPU_MEM_UTIL SERVED_MODEL_NAME \
-    DECODE_CTRL_PORT DECODE_HTTP_PORT PREFILL_PORT ROUTER_PORT \
-    DECODE_WAIT PREFILL_WAIT ROUTER_WAIT SKIP_CONTAINER_BARRIER \
-    CONTAINER_BARRIER_TIMEOUT
-
-LOG_DIR="/run_logs/slurm_job-${SLURM_JOB_ID}"
-SHARED_LOG_DIR="${BENCHMARK_LOGS_DIR}/logs/slurm_job-${SLURM_JOB_ID}"
-mkdir -p "$LOG_DIR"
-
-if [[ "$xP" -ne 1 || "$yD" -ne 1 ]]; then
-    echo "ERROR: tilert supports exactly 1 prefill + 1 decode worker (got xP=$xP yD=$yD)" >&2
-    exit 1
-fi
-if [[ "$NODE_RANK" -lt "$xP" ]]; then
-    TILERT_ROLE=prefill
-else
-    TILERT_ROLE=decode
-fi
-export TILERT_ROLE
-
-source "$WS_PATH/setup_deps.sh"
-source "$WS_PATH/env.sh"
-# benchmark_lib.sh derives AIPERF_DIR from this at source time, so
-# it must be set before the library is loaded, not in run_agentic_replay. The
-# AgentX replay runs in this container, where the repo is mounted at /workspace.
-export INFMAX_CONTAINER_WORKSPACE=/workspace
-source /workspace/benchmarks/benchmark_lib.sh
-
-# Model-specific engine environment (not caller configuration): the prefill
-# vLLM env block lives with the model. Everything else is passed in by the recipe.
-MODELS_YAML="${WS_PATH}/models_tilert.yaml"
-eval "$("$PY" - "$MODELS_YAML" "$MODEL_NAME" <<'PYEOF'
-import shlex, sys, yaml
-path, name = sys.argv[1], sys.argv[2]
-with open(path) as f:
-    models = yaml.safe_load(f) or {}
-if name not in models:
-    sys.exit(f"model '{name}' is not present in {path}")
-m = models[name] or {}
-for key, var in (("prefill_env", "TILERT_PREFILL_ENV"),
-                 ("prefill_extra_flags", "TILERT_PREFILL_EXTRA_FLAGS"),
-                 ("decode_extra_flags", "TILERT_DECODE_EXTRA_FLAGS")):
-    print(f"{var}={shlex.quote(str(m.get(key) or ''))}")
-PYEOF
-)" || { echo "ERROR: cannot read the tilert model entry for '$MODEL_NAME' from $MODELS_YAML" >&2; exit 1; }
-echo "[tilert] model entry '$MODEL_NAME' loaded from $MODELS_YAML"
-
-export ROUTER_PORT
-export SERVED_MODEL_NAME
-
-PREFILL_SPEC=()
-DECODE_MTP=()
-if [[ "$SPEC_DECODING" == "mtp" ]]; then
-    # The prefill rank only has to build the MTP layer's KV; TileRT decode owns
-    # the draft depth (DECODE_MTP_SIZE), so the two counts differ by design.
-    PREFILL_SPEC=(--speculative-config "{\"method\":\"mtp\",\"num_speculative_tokens\":${PREFILL_SPEC_TOKENS}}")
-    # decode_server only accepts depth 3 today, but pass it explicitly so the
-    # converter's --num_mtp, the golden-curve key and the engine depth agree by
-    # data flow rather than by coincidence of defaults.
-    DECODE_MTP=(--with-mtp --num-mtp "$DECODE_MTP_SIZE")
-fi
-
-# The only TileRT recipe on this cluster is AgentX (agentic-coding).
-if [[ "${IS_AGENTIC:-0}" != "1" && "${IS_AGENTIC:-}" != "true" && "${SCENARIO_TYPE:-}" != "agentic-coding" ]]; then
-    echo "ERROR: server_tilert.sh only runs agentic-coding (IS_AGENTIC=${IS_AGENTIC:-} SCENARIO_TYPE=${SCENARIO_TYPE:-})" >&2
-    exit 1
-fi
-
-IFS=',' read -ra IP_ARRAY <<< "$IPADDRS"
-PREFILL_HOST="${IP_ARRAY[0]:-$NODE0_ADDR}"
-DECODE_HOST="${IP_ARRAY[$xP]:-}"
-if [[ -z "$DECODE_HOST" ]]; then
-    echo "ERROR: cannot resolve the decode node IP from IPADDRS='$IPADDRS' (xP=$xP)" >&2
-    exit 1
-fi
-host_ip=$(ip route get 1.1.1.1 2>/dev/null | awk '/src/ {print $7}')
-host_name=$(hostname)
-
-echo "[tilert] ROLE=$TILERT_ROLE rank=$NODE_RANK host=$host_name ($host_ip)"
-echo "[tilert] PREFILL_HOST=$PREFILL_HOST:$PREFILL_PORT  DECODE_HOST=$DECODE_HOST:$DECODE_CTRL_PORT/$DECODE_HTTP_PORT  ROUTER=:$ROUTER_PORT"
-echo "[tilert] MODEL_PATH=$MODEL_PATH  profile=$TILERT_PROFILE  served=$SERVED_MODEL_NAME  max_len=$TILERT_MAX_MODEL_LEN  transport=$TILERT_TRANSPORT  kv=${PREFILL_KV_DTYPE}->${DECODE_KV_DTYPE}  mtp=${SPEC_DECODING}"
-
-# Enable libibverbs fork safety on both ranks before any verbs context exists.
-# Without it, ibv_fork_init() can fail in these containers while Mooncake
-# initialization still reports success, silently falling back from RDMA to TCP.
-# Slower prefill-to-decode KV transfer degrades TTFT; TPOT is unaffected.
-export RDMAV_FORK_SAFE=1
-
-for env_pair in ${TILERT_EXTRA_ENV}; do
-    export "${env_pair?}"
-    echo "[tilert][EXTRA_ENV] $env_pair"
-done
-
-log_and_run_bg() {
-    local label="$1" logfile="$2"; shift 2
-    { printf '===== [%s] %s =====\n' "$label" "$(date '+%F %T')"
-      printf '[cmd]'; printf ' %q' "$@"; printf '\n'
-      printf '[cwd] %s\n[host] %s\n\n' "$PWD" "$host_name"
-    } | tee -a "$logfile"
-    if [[ "$DRY_RUN" -eq 1 ]]; then
-        echo "DRY RUN: [$label] not started"
-        LAST_BG_PID=""
-        return 0
-    fi
-    "$@" >>"$logfile" 2>&1 &
-    LAST_BG_PID=$!
-    echo "[$label] pid=$LAST_BG_PID log=$logfile"
-}
-
-rdma_preflight() {
-    local warn=0
-    echo "[rdma] role=$TILERT_ROLE IBDEVICES=${IBDEVICES:-<unset>} NCCL_SOCKET_IFNAME=${NCCL_SOCKET_IFNAME:-<unset>}"
-    local uverbs=(/dev/infiniband/uverbs*)
-    if [[ -e "${uverbs[0]}" ]]; then
-        echo "[rdma] verbs devices: ${uverbs[*]}"
-    else
-        echo "[rdma] WARNING: /dev/infiniband/uverbs* missing -- the container has no RDMA device nodes (job.slurm passes --device /dev/infiniband)" >&2
-        warn=1
-    fi
-    local ml; ml="$(ulimit -l 2>/dev/null)"
-    if [[ "$ml" == "unlimited" ]]; then
-        echo "[rdma] memlock: unlimited"
-    else
-        echo "[rdma] WARNING: memlock=$ml (not unlimited) -- pinning memory for RDMA may fail (job.slurm passes --ulimit memlock=-1)" >&2
-        warn=1
-    fi
-    if command -v ibv_devices >/dev/null 2>&1; then
-        echo "[rdma] ibv_devices:"; ibv_devices 2>&1 | sed 's/^/[rdma]   /'
-    fi
-    if (( warn )) && [[ "$TILERT_RDMA_STRICT" == "1" ]]; then
-        echo "[rdma] TILERT_RDMA_STRICT=1 and preflight did not fully pass -- aborting" >&2
-        return 1
-    fi
-    return 0
-}
-
-stage_tokenizer_files() {
-    local staged=0 f b
-    for f in "$MODEL_PATH"/*; do
-        [[ -f "$f" ]] || continue
-        b="$(basename "$f")"
-        [[ "$b" == *.safetensors ]] && continue
-        [[ "$b" == "model.safetensors.index.json" ]] && continue
-        [[ -e "$TILERT_WEIGHTS_DIR/$b" ]] && continue
-        cp -p "$f" "$TILERT_WEIGHTS_DIR/$b" && staged=$((staged+1))
-    done
-    echo "[stage_tokenizer] staged $staged auxiliary file(s) from $MODEL_PATH -> $TILERT_WEIGHTS_DIR"
-    local missing=()
-    [[ -f "$TILERT_WEIGHTS_DIR/chat_template.jinja" ]] || missing+=(chat_template.jinja)
-    [[ -f "$TILERT_WEIGHTS_DIR/tokenizer_config.json" || -f "$TILERT_WEIGHTS_DIR/tokenizer.json" ]] \
-        || missing+=("tokenizer.json/tokenizer_config.json")
-    if (( ${#missing[@]} )); then
-        echo "[stage_tokenizer] ERROR: $TILERT_WEIGHTS_DIR is missing ${missing[*]}; check that MODEL_PATH=$MODEL_PATH is an HF directory with the tokenizer" >&2
-        return 1
-    fi
-    return 0
-}
-
-_tilert_weights_cached() {
-    local r
-    for r in $(seq 0 $((DECODE_TP_SIZE - 1))); do
-        [[ -f "$TILERT_WEIGHTS_DIR/rank${r}/model.safetensors.index.json" ]] || return 1
-    done
-    # The engine refuses a cache converted without the MTP module (end2end.py
-    # checks tilert_meta.json num_mtp), but only after loading ~90 GiB of
-    # weights. Check the same field here so a stale non-MTP cache is
-    # re-converted instead of failing late.
-    [[ -f "$TILERT_WEIGHTS_DIR/tilert_meta.json" ]] || return 1
-    if [[ "$SPEC_DECODING" == "mtp" ]]; then
-        "$PY" - "$TILERT_WEIGHTS_DIR/tilert_meta.json" <<'PYEOF' || return 1
-import json, sys
-sys.exit(0 if int(json.load(open(sys.argv[1])).get("num_mtp", 0)) >= 1 else 1)
-PYEOF
-    fi
-    return 0
-}
-
-convert_weights() {
-    if _tilert_weights_cached; then
-        echo "[weight_converter] cache hit (${DECODE_TP_SIZE}/${DECODE_TP_SIZE} rank index.json), skipping conversion: $TILERT_WEIGHTS_DIR"
-        return 0
-    fi
-    mkdir -p "$TILERT_WEIGHTS_DIR" || { echo "[weight_converter] ERROR: cannot create $TILERT_WEIGHTS_DIR (set TILERT_WEIGHTS_DIR to a writable shared path)" >&2; return 1; }
-    exec 9>"$TILERT_WEIGHTS_DIR/.convert.lock"
-    flock -w "$TILERT_CONVERT_LOCK_WAIT" 9 || {
-        echo "[weight_converter] timed out waiting for the conversion lock (another job still converting?)" >&2; return 1; }
-    if _tilert_weights_cached; then
-        echo "[weight_converter] cache produced by a concurrent job, skipping conversion"; exec 9>&-; return 0
-    fi
-    if [[ -n "$(ls -A "$TILERT_WEIGHTS_DIR" 2>/dev/null | grep -v '^\.convert\.lock$')" ]]; then
-        echo "[weight_converter] leftovers without index.json (previous conversion incomplete); cleaning and re-converting"
-        find "$TILERT_WEIGHTS_DIR" -mindepth 1 ! -name '.convert.lock' -delete
-    fi
-    echo "[weight_converter] $MODEL_PATH -> $TILERT_WEIGHTS_DIR (model_type=$TILERT_MODEL_TYPE)"
-    local conv_mod conv_args
-    if "$PY" -c "import tilert.models.${TILERT_MODEL_PKG}.weight_converter" 2>/dev/null; then
-        conv_mod="tilert.models.${TILERT_MODEL_PKG}.weight_converter"
-        conv_args=(--model_dir "$MODEL_PATH" --save_dir "$TILERT_WEIGHTS_DIR"
-                   --device "cuda:$((GPUS_PER_NODE - 1))")
-        [[ "$SPEC_DECODING" == "mtp" ]] && conv_args+=(--num_mtp "$DECODE_MTP_SIZE")
-    else
-        conv_mod="tilert.models.preprocess.weight_converter"
-        conv_args=(--model_type "$TILERT_MODEL_TYPE" --model_dir "$MODEL_PATH" --save_dir "$TILERT_WEIGHTS_DIR")
-    fi
-    if [[ "$DRY_RUN" -eq 1 ]]; then
-        echo "DRY RUN: $PY -m $conv_mod ${conv_args[*]}"
-        exec 9>&-; return 0
-    fi
-    echo "[weight_converter] using $conv_mod"
-    "$PY" -m "$conv_mod" "${conv_args[@]}" \
-        2>&1 | tee "$LOG_DIR/tilert_weight_converter_${host_name}.log"
-    local rc=${PIPESTATUS[0]}
-    exec 9>&-
-    if [[ $rc -ne 0 ]] || ! _tilert_weights_cached; then
-        echo "[weight_converter] ERROR: conversion failed (rc=$rc, per-rank index.json complete: $(_tilert_weights_cached && echo yes || echo no))" >&2
-        return 1
-    fi
-    echo "[weight_converter] conversion done and cached: $TILERT_WEIGHTS_DIR"
-}
-
-start_decode() {
-    # shellcheck disable=SC2206
-    local extra=( ${TILERT_DECODE_EXTRA_FLAGS} )
-    if [[ "$SPEC_DECODING" == "mtp" && "$EVAL_ONLY" != "true" && "$RUN_EVAL" != "true" ]]; then
-        check_env_vars MODEL_PREFIX THINKING_MODE
-        local curve="${WS_PATH%/benchmarks/*}/infx/golden_al_distribution/${MODEL_PREFIX}_mtp.yaml"
-        TILERT_SIMULATE_ACC_LEN="$("$PY" - "$curve" "$THINKING_MODE" "$DECODE_MTP_SIZE" <<'PYEOF'
-import sys, yaml
-path, thinking, tokens = sys.argv[1], sys.argv[2], int(sys.argv[3])
-data = yaml.safe_load(open(path))
-if not isinstance(data, dict) or len(data) != 1:
-    sys.exit(f"golden curve {path} must hold exactly one model key")
-model, modes = next(iter(data.items()))
-try:
-    value = float(modes[thinking][tokens])
-except (KeyError, TypeError, ValueError):
-    sys.exit(f"no golden acceptance for {model}/{thinking}/{tokens} draft tokens in {path}")
-if not 1 <= value <= tokens + 1:
-    sys.exit(f"golden acceptance {value} out of range for {tokens} draft tokens")
-print(f"{value:g}")
-PYEOF
-)" || { echo "[tilert] ERROR: golden AL lookup failed (curve=$curve)" >&2; exit 1; }
-        echo "[tilert] golden AL ${TILERT_SIMULATE_ACC_LEN} from $(basename "$curve") ($THINKING_MODE, K=${DECODE_MTP_SIZE})"
-    fi
-
-    if [[ -n "${TILERT_SIMULATE_ACC_LEN:-}" && "$EVAL_ONLY" != "true" ]]; then
-        export TILERT_SIMULATE_ACC_LEN
-        export TILERT_SIMULATE_ACC_METHOD
-        echo "[decode] simulated acceptance: TILERT_SIMULATE_ACC_LEN=${TILERT_SIMULATE_ACC_LEN}" \
-             "method=${TILERT_SIMULATE_ACC_METHOD} (output text is meaningless by design)"
-    else
-        unset TILERT_SIMULATE_ACC_LEN TILERT_SIMULATE_ACC_METHOD
-        echo "[decode] real MTP verification (no simulated acceptance)"
-    fi
-    local cmd=("$PY" -m tilert.pd_vllm.decode_server
-        --engine tilert --model "$TILERT_PROFILE"
-        --model-weights-dir "$TILERT_WEIGHTS_DIR"
-        --max-seq-len "$TILERT_MAX_MODEL_LEN"
-        --kv-cache-dtype "$DECODE_KV_DTYPE" --transport "$TILERT_TRANSPORT"
-        --ctrl-port "$DECODE_CTRL_PORT" --http-port "$DECODE_HTTP_PORT"
-        "${DECODE_MTP[@]}" "${extra[@]}")
-    log_and_run_bg decode "$LOG_DIR/decode_${host_name}.log" "${cmd[@]}"
-    DECODE_PID=$LAST_BG_PID
-}
-
-start_prefill() {
-    for env_pair in ${TILERT_PREFILL_ENV}; do
-        export "${env_pair?}"
-        echo "[PREFILL_ENV] $env_pair"
-    done
-    local served=("$SERVED_MODEL_NAME")
-    [[ -n "$MODEL_NAME" && "$MODEL_NAME" != "$SERVED_MODEL_NAME" ]] && served+=("$MODEL_NAME")
-    # shellcheck disable=SC2206
-    local extra=( ${TILERT_PREFILL_EXTRA_FLAGS} )
-    local kv_cfg
-    kv_cfg=$(printf '{"kv_connector":"TileRTConnector","kv_connector_module_path":"tilert.pd_vllm.prefill_connector","kv_role":"kv_producer","kv_connector_extra_config":{"tilert_host":"%s","tilert_ctrl_port":%s,"tilert_model":"%s","tilert_max_seq_len":%s,"tilert_transport":"%s"}}' \
-        "$DECODE_HOST" "$DECODE_CTRL_PORT" "$TILERT_PROFILE" "$TILERT_MAX_MODEL_LEN" "$TILERT_TRANSPORT")
-    local cmd=(vllm serve "$MODEL_PATH"
-        --served-model-name "${served[@]}" --port "$PREFILL_PORT"
-        --tensor-parallel-size "$PREFILL_TP_SIZE" --max-model-len "$TILERT_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"
-        --block-size "$PREFILL_BLOCK_SIZE"
-        "${PREFILL_SPEC[@]}"
-        --kv-transfer-config "$kv_cfg"
-        "${extra[@]}")
-    log_and_run_bg prefill "$LOG_DIR/prefill_${host_name}.log" "${cmd[@]}"
-    PREFILL_PID=$LAST_BG_PID
-}
-
-start_router() {
-    local cmd=(env HIP_VISIBLE_DEVICES= ROCR_VISIBLE_DEVICES= CUDA_VISIBLE_DEVICES=
-        "$PY" -m tilert.pd_vllm.pd_router
-        --vllm-url "http://$PREFILL_HOST:$PREFILL_PORT"
-        --decode "$DECODE_HOST:$DECODE_CTRL_PORT:$DECODE_HTTP_PORT"
-        --host 0.0.0.0 --port "$ROUTER_PORT" --model-path "$MODEL_PATH" --parser "$TILERT_PARSER"
-        --queue-timeout "$TILERT_QUEUE_TIMEOUT")
-    log_and_run_bg router "$LOG_DIR/router_${host_name}.log" "${cmd[@]}"
-    ROUTER_PID=$LAST_BG_PID
-}
-
-tcp_open() { (exec 3<>"/dev/tcp/$1/$2") 2>/dev/null; }
-
-wait_for_tcp() {
-    local host="$1" port="$2" timeout="${3:-600}" pid="${4:-}"
-    local deadline=$(( SECONDS + timeout ))
-    until tcp_open "$host" "$port"; do
-        if [[ -n "$pid" ]] && ! kill -0 "$pid" 2>/dev/null; then
-            echo "[wait_for_tcp] process $pid exited before $host:$port opened" >&2; return 2
-        fi
-        if [[ $SECONDS -ge $deadline ]]; then
-            echo "[wait_for_tcp] timeout: $host:$port not open after ${timeout}s" >&2; return 1
-        fi
-        sleep 5
-    done
-    echo "[wait_for_tcp] $host:$port ready"
-}
-
-wait_for_tcp_close() {
-    local host="$1" port="$2" pid="${3:-}"
-    while tcp_open "$host" "$port"; do
-        if [[ -n "$pid" ]] && ! kill -0 "$pid" 2>/dev/null; then
-            echo "[wait_for_tcp_close] process $pid exited while $host:$port is still open" >&2; return 2
-        fi
-        sleep 10
-    done
-    echo "[wait_for_tcp_close] $host:$port closed"
-}
-
-copy_logs_to_shared() {
-    [[ "$DRY_RUN" -eq 0 ]] || return 0
-    mkdir -p "$SHARED_LOG_DIR" && cp -r "$LOG_DIR"/. "$SHARED_LOG_DIR"/ \
-        && echo "Copied $LOG_DIR -> $SHARED_LOG_DIR" \
-        || echo "WARNING: failed to copy $LOG_DIR to $SHARED_LOG_DIR" >&2
-}
-trap copy_logs_to_shared EXIT
-
-run_lm_eval_on_router() {
-    echo "Running lm-eval evaluation on the router..."
-    local ok=false _attempt
-    for _attempt in 1 2 3; do
-        if curl -sf --max-time 10 "http://0.0.0.0:${ROUTER_PORT}/health" >/dev/null 2>&1; then ok=true; break; fi
-        echo "Eval health check attempt $_attempt failed, retrying in 10s..."; sleep 10
-    done
-    if [[ "$ok" != "true" ]]; then
-        echo "ERROR: router health check failed after 3 attempts; skipping eval" >&2
-        return 1
-    fi
-    local eval_failed=0
-    pushd /workspace >/dev/null || return 1
-    if [[ -n "${EVAL_CONC:-}" ]]; then
-        export EVAL_CONCURRENT_REQUESTS="${EVAL_CONC}"
-    else
-        export EVAL_CONCURRENT_REQUESTS=$(echo "$BENCH_MAX_CONCURRENCY" | tr 'x' '\n' | sort -n | tail -1)
-    fi
-    # run_lm_eval reads the endpoint from PORT (check_env_vars) and names the
-    # model from MODEL_NAME, which the vLLM prefill also serves next to
-    # SERVED_MODEL_NAME; MODEL stays the local HF dir for the context lookup.
-    export PORT="$ROUTER_PORT"
-    export MODEL="$MODEL_PATH"
-    export MAX_MODEL_LEN="$TILERT_MAX_MODEL_LEN"
-    if [[ "$DRY_RUN" -eq 1 ]]; then
-        echo "DRY RUN: run_eval --port $ROUTER_PORT (framework=${EVAL_FRAMEWORK}, conc=${EVAL_CONCURRENT_REQUESTS})"
-    else
-        run_eval --port "$ROUTER_PORT"
-        local eval_rc=$?
-        if [[ $eval_rc -ne 0 ]]; then
-            echo "ERROR: run_eval exited rc=$eval_rc; preserving failure artifacts" >&2
-            eval_failed=1
-        else
-            export TP="${PREFILL_TP_SIZE}" CONC="${EVAL_CONCURRENT_REQUESTS}" EP_SIZE=1
-            export PREFILL_TP="${PREFILL_TP_SIZE}" PREFILL_EP=1 PREFILL_NUM_WORKERS="${xP}"
-            export DECODE_TP="${DECODE_TP_SIZE}" DECODE_EP=1 DECODE_NUM_WORKERS="${yD}"
-            export DP_ATTENTION=false PREFILL_DP_ATTENTION=false DECODE_DP_ATTENTION=false
-            export ISL="${BENCH_INPUT_LEN}" OSL="${BENCH_OUTPUT_LEN}"
-            # As on the SGLang path: rewrite meta_env.json from the exports above,
-            # then stage unless run_eval already did (eval-only).
-            rewrite_lm_eval_meta_env
-            if [[ "$EVAL_ONLY" != "true" ]]; then
-                append_lm_eval_summary
-            fi
-        fi
-        local eval_copy_dir="$LOG_DIR/eval_results"
-        if stage_eval_artifacts "$eval_copy_dir" /workspace "${EVAL_RESULT_DIR:-}"; then
-            echo "Eval artifacts staged in $eval_copy_dir"
-        else
-            echo "ERROR: failed to stage eval artifacts in $eval_copy_dir" >&2
-            eval_failed=1
-        fi
-    fi
-    popd >/dev/null || true
-    return $eval_failed
-}
-
-run_agentic_replay() {
-    local rc=0
-    wait_for_server_ready --port "$ROUTER_PORT" --server-log "$LOG_DIR/router_${host_name}.log" --server-pid "$ROUTER_PID"
-    cd /workspace || return 1
-
-    export PORT="$ROUTER_PORT"
-    export MODEL="$MODEL_PATH"              # aiperf --tokenizer (local HF dir)
-    export SERVED_MODEL_NAME                # aiperf --model (name the router/vLLM serve)
-    check_env_vars DURATION RESULT_FILENAME INFMAX_CONTAINER_WORKSPACE
-    export MAX_MODEL_LEN="$TILERT_MAX_MODEL_LEN"
-    # TileRT decode exposes no /metrics route; only the vLLM prefill is scraped.
-    export AIPERF_SERVER_METRICS_URLS="http://${PREFILL_HOST}:${PREFILL_PORT}/metrics"
-    export TRANSFORMERS_VERBOSITY=error TOKENIZERS_PARALLELISM=false
-    # Keep the trace corpus and aiperf's HF downloads on the node's /tmp mount
-    # instead of the container's ephemeral ~/.cache, as the SGLang client does.
-    export HF_HOME=/run_logs/hf_cache
-
-    local result_dir="$LOG_DIR/agentic"
-    local result_filename_base="$RESULT_FILENAME"
-    mkdir -p "$result_dir"
-
-    # Neither server.sh nor this script runs with errexit; a failed bootstrap
-    # must not fall through into the replay loop and its misleading cascade.
-    resolve_trace_source || return 1
-    install_agentic_deps || return 1
-
-    local conc conc_result_dir
-    for conc in ${BENCH_MAX_CONCURRENCY//x/ }; do
-        echo "=========================================="
-        echo "Agentic trace replay: conc=$conc"
-        echo "=========================================="
-        conc_result_dir="$result_dir/conc_${conc}"
-        mkdir -p "$conc_result_dir"
-        export CONC="$conc" USERS="$conc"
-        build_replay_cmd "$conc_result_dir"
-        export RESULT_FILENAME="${result_filename_base}_conc${conc}"
-        if [[ "$DRY_RUN" -eq 1 ]]; then
-            echo "DRY RUN: $REPLAY_CMD"
-        elif ! run_agentic_replay_and_write_outputs "$conc_result_dir"; then
-            echo "WARNING: agentic trace replay for conc=$conc failed (replay or validation) after writing available results" >&2
-            rc=1
-        fi
-        echo "-----------------------------------------"
-    done
-    export RESULT_FILENAME="$result_filename_base"
-    return $rc
-}
-
-echo "Waiting at the container creation barrier on $host_name"
-if [[ "$DRY_RUN" -eq 1 ]]; then
-    echo "DRY RUN: skipping container creation barrier"
-elif [[ "$SKIP_CONTAINER_BARRIER" == "1" ]]; then
-    echo "SKIP_CONTAINER_BARRIER=1: caller asserts all containers are up"
-else
-    # --grace 60: after the barrier passes, sync.py keeps the port open for
-    # max(60, timeout/2) seconds in the foreground so a peer one poll behind
-    # still sees it. At CONTAINER_BARRIER_TIMEOUT=5400 that is a 45-minute idle
-    # sleep on every rank (jobs 45373/45374 slept 06:28-07:13). Both ranks pass
-    # within one 5 s poll of each other and the stages below have their own
-    # readiness waits, so 60 s is plenty.
-    "$PY" "$WS_PATH/sync.py" barrier \
-        --local-ip "${host_ip}" --local-port 5000 --enable-port \
-        --node-ips "${IPADDRS}" --node-ports 5000 \
-        --wait-for-all-ports --timeout "$CONTAINER_BARRIER_TIMEOUT" --grace 60 \
-        || { echo "ERROR: container creation barrier failed after ${CONTAINER_BARRIER_TIMEOUT}s -- the peer rank never opened port 5000." \
-                  "A cold image pull is the usual cause: this recipe pulls two ~32 GB images, one per rank, and the rank that" \
-                  "comes up first waits out the whole timeout while the other is still pulling." >&2; exit 1; }
-fi
-
-case "$TILERT_ROLE" in
-    decode)
-        echo "${host_name}:${host_ip} is the TileRT Decode Node (Model: ${MODEL_NAME}, profile: ${TILERT_PROFILE})"
-        rdma_preflight || exit 1
-        convert_weights || exit 1
-        stage_tokenizer_files || exit 1
-        start_decode
-        if [[ "$DRY_RUN" -eq 1 ]]; then
-            echo "DRY RUN: decode role complete"; exit 0
-        fi
-        echo "Waiting for the router port ${PREFILL_HOST}:${ROUTER_PORT} to open (timeout ${ROUTER_WAIT}s)..."
-        wait_for_tcp "$PREFILL_HOST" "$ROUTER_PORT" "$ROUTER_WAIT" "$DECODE_PID"; wrc=$?
-        if [[ $wrc -eq 2 ]]; then
-            echo "ERROR: decode_server exited before the router came up (see $LOG_DIR/decode_${host_name}.log)" >&2
-            tail -50 "$LOG_DIR/decode_${host_name}.log" >&2 || true
-            copy_logs_to_shared; exit 1
-        elif [[ $wrc -ne 0 ]]; then
-            echo "WARNING: router never opened within ${ROUTER_WAIT}s; shutting down decode" >&2
-            kill "$DECODE_PID" 2>/dev/null || true
-            copy_logs_to_shared; exit 1
-        fi
-        echo "Waiting until the router port closes..."
-        wait_for_tcp_close "$PREFILL_HOST" "$ROUTER_PORT" "$DECODE_PID"; wrc=$?
-        if [[ $wrc -eq 2 ]]; then
-            echo "ERROR: decode_server died while the benchmark was running (see $LOG_DIR/decode_${host_name}.log)" >&2
-            tail -50 "$LOG_DIR/decode_${host_name}.log" >&2 || true
-            copy_logs_to_shared; exit 1
-        fi
-        echo "Killing the decode server"
-        kill "$DECODE_PID" 2>/dev/null || true
-        sleep 2
-        copy_logs_to_shared
-        ;;
-    prefill)
-        echo "NODE INFO ======================================="
-        echo "Node List : ${SLURM_JOB_NODELIST:-}"
-        echo "Node IPs  : ${IPADDRS}"
-        echo "Model     : ${MODEL_NAME}"
-        echo "${host_name}:${host_ip} is the Prefill Node (vLLM + TileRTConnector) and Router Node"
-        echo "================================================"
-        rdma_preflight || exit 1
-        echo "Waiting for the decode ctrl port ${DECODE_HOST}:${DECODE_CTRL_PORT} (timeout ${DECODE_WAIT}s)..."
-        if [[ "$DRY_RUN" -eq 0 ]]; then
-            wait_for_tcp "$DECODE_HOST" "$DECODE_CTRL_PORT" "$DECODE_WAIT" \
-                || echo "WARNING: timed out waiting for the decode ctrl port; starting prefill anyway" >&2
-        fi
-        start_prefill
-        if [[ "$DRY_RUN" -eq 0 ]]; then
-            wait_for_tcp "$PREFILL_HOST" "$PREFILL_PORT" "$PREFILL_WAIT" "$PREFILL_PID"; wrc=$?
-            if [[ $wrc -ne 0 ]]; then
-                echo "ERROR: vLLM prefill did not open ${PREFILL_HOST}:${PREFILL_PORT} (rc=$wrc, see $LOG_DIR/prefill_${host_name}.log)" >&2
-                tail -50 "$LOG_DIR/prefill_${host_name}.log" >&2 || true
-                kill "$PREFILL_PID" 2>/dev/null || true
-                copy_logs_to_shared; exit 1
-            fi
-        fi
-        start_router
-        if [[ "$DRY_RUN" -eq 1 ]]; then
-            echo "DRY RUN: prefill/router role complete"; exit 0
-        fi
-        echo "Ready for benchmarking on ${host_name}:${host_ip}"
-        cd "$WS_PATH" || exit 1
-        # EVAL_ONLY skips the AgentX replay and runs GSM8K on the same router;
-        # RUN_EVAL after a replay runs it once the replay has finished.
-        if [[ "$EVAL_ONLY" == "true" ]]; then
-            echo "EVAL_ONLY mode: skipping the AgentX replay"
-            wait_for_server_ready --port "$ROUTER_PORT" --server-log "$LOG_DIR/router_${host_name}.log" --server-pid "$ROUTER_PID"
-            export TRANSFORMERS_VERBOSITY=error TOKENIZERS_PARALLELISM=false
-            run_lm_eval_on_router; BENCH_RC=$?
-        else
-            run_agentic_replay; BENCH_RC=$?
-            if [[ "$RUN_EVAL" == "true" ]]; then
-                run_lm_eval_on_router || BENCH_RC=1
-            fi
-        fi
-        copy_logs_to_shared
-        echo "Killing the router and the prefill server"
-        kill "$ROUTER_PID" "$PREFILL_PID" 2>/dev/null || true
-        sleep 2
-        pkill -f "tilert.pd_vllm.pd_router" 2>/dev/null || true
-        pkill -f "vllm serve" 2>/dev/null || true
-        if [[ "$BENCH_RC" -ne 0 ]]; then
-            echo "ERROR: benchmark/eval reported rc=$BENCH_RC" >&2
-            exit "$BENCH_RC"
-        fi
-        ;;
-    *)
-        echo "ERROR: unknown TILERT_ROLE='$TILERT_ROLE'" >&2; exit 2 ;;
-esac
-
-echo "Script completed successfully"
-exit 0
diff --git a/inferencex-e2e/benchmarks/multi_node/amd_utils/setup_deps.sh b/inferencex-e2e/benchmarks/multi_node/amd_utils/setup_deps.sh
deleted file mode 100644
index 1a5c5cfacf..0000000000
--- a/inferencex-e2e/benchmarks/multi_node/amd_utils/setup_deps.sh
+++ /dev/null
@@ -1,136 +0,0 @@
-#!/bin/bash
-
-source "$(dirname "${BASH_SOURCE[0]}")/../../benchmark_lib.sh" --validation-only
-# Install missing disagg dependencies at container start. Each installer is
-# idempotent and gated on $ENGINE.
-
-_SETUP_START=$(date +%s)
-_SETUP_INSTALLED=()
-
-# Pinned by the recipe (TILERT_VERSION); the rest are fixed properties of the
-# TileRT 0.1.x runtime rather than caller configuration.
-TILERT_PACKAGE=tilert
-TILERT_HTTP_DEPS="fastapi uvicorn httpx"
-TILERT_TRANSPORT_DEPS="mooncake-transfer-engine-rocm>=0.3.13"
-TILERT_TRANSFORMERS_SPEC="transformers>=4.56"
-
-_tilert_resolve_python() {
-    if [[ -n "${PY:-}" ]] && command -v "$PY" >/dev/null 2>&1; then :; else
-        PY=""
-        local c
-        for c in python3 python; do command -v "$c" >/dev/null 2>&1 && { PY="$c"; break; }; done
-    fi
-    [[ -n "$PY" ]] || { echo "[SETUP] ERROR: neither python3 nor python found"; exit 1; }
-    export PY
-    echo "[SETUP] interpreter PY=$PY ($(command -v "$PY"))"
-}
-
-_tilert_installed_version() {
-    "$PY" - "$1" <<'PYEOF' 2>/dev/null
-import sys
-from importlib.metadata import version, PackageNotFoundError
-try:
-    print(version(sys.argv[1]))
-except PackageNotFoundError:
-    pass
-PYEOF
-}
-
-_tilert_pip() {
-    "$PY" -m pip install --quiet --no-cache-dir "$@"
-}
-
-_tilert_install_missing() {
-    local probe="$1"; shift
-    [[ $# -gt 0 ]] || return 0
-    if "$PY" -c "import $probe" 2>/dev/null; then
-        echo "[SETUP] $probe already present, skipping ($*)"
-        return 0
-    fi
-    echo "[SETUP] installing $* (probe module '$probe' missing)"
-    _tilert_pip "$@" || { echo "[SETUP] ERROR: failed to install: $*"; exit 1; }
-    _SETUP_INSTALLED+=("$*")
-}
-
-install_tilert_container_tools() {
-    if command -v ip >/dev/null 2>&1 && command -v curl >/dev/null 2>&1 \
-        && command -v ibv_devices >/dev/null 2>&1; then
-        echo "[SETUP] Container RDMA/net tools already present"
-        return 0
-    fi
-    echo "[SETUP] Installing iproute2 + curl + ibverbs userspace in container..."
-    apt-get update -q -y && apt-get install -q -y --no-install-recommends \
-        iproute2 curl ibverbs-utils libibverbs1 librdmacm1 ibverbs-providers \
-        && rm -rf /var/lib/apt/lists/*
-    if ! command -v ip >/dev/null 2>&1 || ! command -v curl >/dev/null 2>&1; then
-        echo "[SETUP] ERROR: failed to install iproute2/curl"; exit 1
-    fi
-    _SETUP_INSTALLED+=("iproute2+curl+ibverbs")
-}
-
-_tilert_install_wheel() {
-    local mode="$1"  # full | no-deps
-    local have; have="$(_tilert_installed_version tilert)"
-    if [[ "$have" == "$TILERT_VERSION" ]]; then
-        echo "[SETUP] tilert $have already installed, skipping"
-        return 0
-    fi
-    [[ -n "$have" ]] && echo "[SETUP] tilert $have installed, switching to pinned $TILERT_VERSION"
-    if [[ "$mode" == "no-deps" ]]; then
-        echo "[SETUP] installing $TILERT_PIP_SPEC --no-deps (connector plugin + router on top of the image's vLLM)"
-        _tilert_pip --no-deps "$TILERT_PIP_SPEC" || { echo "[SETUP] ERROR: failed to install $TILERT_PIP_SPEC (--no-deps)"; exit 1; }
-    else
-        echo "[SETUP] installing $TILERT_PIP_SPEC (TileRT ROCm build, official PyPI wheel)"
-        _tilert_pip "$TILERT_PIP_SPEC" || { echo "[SETUP] ERROR: failed to install $TILERT_PIP_SPEC"; exit 1; }
-    fi
-    have="$(_tilert_installed_version tilert)"
-    [[ "$have" == "$TILERT_VERSION" ]] || {
-        echo "[SETUP] ERROR: tilert is ${have:-not installed} after install, expected $TILERT_VERSION"; exit 1; }
-    _SETUP_INSTALLED+=("$TILERT_PACKAGE==$TILERT_VERSION($mode)")
-}
-
-install_tilert_decode() {
-    install_tilert_container_tools
-    _tilert_install_wheel full
-    _tilert_install_missing uvicorn $TILERT_HTTP_DEPS
-    _tilert_install_missing mooncake.engine "$TILERT_TRANSPORT_DEPS"
-    _tilert_install_missing transformers "$TILERT_TRANSFORMERS_SPEC"
-    "$PY" -c "import tilert.pd_vllm.decode_server" 2>/dev/null || {
-        echo "[SETUP] ERROR: import tilert.pd_vllm.decode_server failed:"
-        "$PY" -c "import tilert.pd_vllm.decode_server" 2>&1 | tail -3
-        exit 1; }
-    echo "[SETUP] tilert.pd_vllm.decode_server imports OK"
-}
-
-install_tilert_prefill() {
-    local vllm_v; vllm_v="$(_tilert_installed_version vllm)"
-    if [[ -z "$vllm_v" ]]; then
-        echo "[SETUP] ERROR: no vLLM in the prefill image (PREFILL_IMAGE must be a vllm/vllm-openai-rocm image)."
-        exit 1
-    fi
-    echo "[SETUP] prefill-side vLLM $vllm_v"
-    install_tilert_container_tools
-    _tilert_install_wheel no-deps
-    _tilert_install_missing mooncake.engine "$TILERT_TRANSPORT_DEPS"
-    "$PY" -c "import tilert.pd_vllm.prefill_connector" 2>/dev/null || {
-        echo "[SETUP] WARN: import tilert.pd_vllm.prefill_connector failed (vLLM will report again when loading the connector plugin):"
-        "$PY" -c "import tilert.pd_vllm.prefill_connector" 2>&1 | tail -3; }
-}
-
-if [[ "$ENGINE" == "tilert" ]]; then
-    check_env_vars TILERT_VERSION
-    TILERT_PIP_SPEC="$TILERT_PACKAGE==$TILERT_VERSION"
-    _tilert_resolve_python
-    case "${TILERT_ROLE:-}" in
-        decode)  install_tilert_decode ;;
-        prefill) install_tilert_prefill ;;
-        *) echo "[SETUP] ERROR: ENGINE=tilert needs TILERT_ROLE=decode|prefill (got '${TILERT_ROLE:-}')"; exit 1 ;;
-    esac
-fi
-
-_SETUP_END=$(date +%s)
-if [[ ${#_SETUP_INSTALLED[@]} -eq 0 ]]; then
-    echo "[SETUP] All dependencies already present ($(( _SETUP_END - _SETUP_START ))s wallclock)"
-else
-    echo "[SETUP] Installed: ${_SETUP_INSTALLED[*]} in $(( _SETUP_END - _SETUP_START ))s"
-fi
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
new file mode 100644
index 0000000000..1c636f8229
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+source /infmax-workspace/benchmarks/benchmark_lib.sh --validation-only
+check_env_vars TILERT_VERSION TILERT_ROLE
+case "$TILERT_ROLE" in
+    prefill|decode|router) ;;
+    *) echo "Unknown TileRT role: $TILERT_ROLE" >&2; exit 1 ;;
+esac
+
+install_args=(--quiet --no-cache-dir)
+if [[ "$TILERT_ROLE" == prefill ]]; then
+    # Preserve the prefill image's vLLM/Torch dependency set.
+    install_args+=(--no-deps)
+fi
+python3 -m pip install "${install_args[@]}" "tilert==$TILERT_VERSION"
+
+if ! python3 -c 'import mooncake.engine' >/dev/null 2>&1; then
+    python3 -m pip install --quiet --no-cache-dir 'mooncake-transfer-engine-rocm>=0.3.13'
+fi
+if [[ "$TILERT_ROLE" != prefill ]]; then
+    if ! python3 -c 'import uvicorn' >/dev/null 2>&1; then
+        python3 -m pip install --quiet --no-cache-dir fastapi uvicorn httpx
+    fi
+    if ! python3 -c 'import transformers' >/dev/null 2>&1; then
+        python3 -m pip install --quiet --no-cache-dir 'transformers>=4.56'
+    fi
+fi
+
+if [[ "$TILERT_ROLE" != decode ]]; then
+    exit 0
+fi
+check_env_vars TILERT_WEIGHTS_DIR TILERT_CONVERT_LOCK_WAIT
+
+weights_ready() {
+    local rank
+    for rank in {0..7}; do
+        [[ -f "$TILERT_WEIGHTS_DIR/rank$rank/model.safetensors.index.json" ]] || return 1
+    done
+    [[ -f "$TILERT_WEIGHTS_DIR/tilert_meta.json" ]] || return 1
+    python3 - "$TILERT_WEIGHTS_DIR/tilert_meta.json" <<'PY'
+import json
+import sys
+
+with open(sys.argv[1]) as source:
+    metadata = json.load(source)
+sys.exit(0 if int(metadata.get("num_mtp", 0)) >= 1 else 1)
+PY
+}
+
+mkdir -p "$TILERT_WEIGHTS_DIR"
+exec 9>"$TILERT_WEIGHTS_DIR/.convert.lock"
+flock -w "$TILERT_CONVERT_LOCK_WAIT" 9
+if ! weights_ready; then
+    python3 -m tilert.models.glm_5_2_rocm.weight_converter \
+        --model_dir /model --save_dir "$TILERT_WEIGHTS_DIR" --device cuda:7 --num_mtp 3
+    weights_ready
+fi
+
+for file in /model/*; do
+    [[ -f "$file" ]] || continue
+    name="${file##*/}"
+    [[ "$name" == *.safetensors || "$name" == model.safetensors.index.json ]] && continue
+    [[ -e "$TILERT_WEIGHTS_DIR/$name" ]] || cp -p "$file" "$TILERT_WEIGHTS_DIR/$name"
+done
+[[ -f "$TILERT_WEIGHTS_DIR/chat_template.jinja" ]]
+[[ -f "$TILERT_WEIGHTS_DIR/tokenizer_config.json" || -f "$TILERT_WEIGHTS_DIR/tokenizer.json" ]]
+echo "TileRT TP8/MTP weights ready: $TILERT_WEIGHTS_DIR"
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
new file mode 100644
index 0000000000..b9215a3759
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
@@ -0,0 +1,103 @@
+schema: 2
+name: glm5.3-mi355x-tilert-agentx
+model:
+  path: GLM-5.3
+  container: ghcr.io/tile-ai/tilert-rocm-decode:0.1.6
+  precision: fp8
+slurm:
+  time_limit: "08:00:00"
+resources:
+  gpu_type: mi355x
+  gpus_per_node: 8
+setup_script: glm5.3-tilert-rocm.sh
+environment:
+  TILERT_VERSION: "0.1.6.post2"
+  RDMAV_FORK_SAFE: "1"
+  PYTHONDONTWRITEBYTECODE: "1"
+  NCCL_SOCKET_IFNAME: eno0
+  GLOO_SOCKET_IFNAME: eno0
+  NCCL_IB_HCA: rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7
+roles:
+  prefill:
+    engine:
+      type: vllm
+      set_visible_devices: true
+    container: ghcr.io/tile-ai/tilert-rocm-prefill:0.1.6
+    nodes: 1
+    workers: 1
+    gpus: 8
+    env:
+      TILERT_ROLE: prefill
+      VLLM_ROCM_USE_AITER: "1"
+      VLLM_ROCM_USE_AITER_MOE: "1"
+      VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1"
+      VLLM_ENGINE_READY_TIMEOUT_S: "10800"
+    args:
+      served-model-name: [glm5_2, GLM-5.3]
+      tensor-parallel-size: 8
+      max-model-len: 1048576
+      enforce-eager: true
+      trust-remote-code: true
+      return-tokens-as-token-ids: true
+      gpu-memory-utilization: 0.85
+      kv-cache-dtype: bfloat16
+      block-size: 64
+      speculative-config: '{"method":"mtp","num_speculative_tokens":1}'
+      kv-transfer-config: >-
+        {"kv_connector":"TileRTConnector","kv_connector_module_path":"tilert.pd_vllm.prefill_connector",
+         "kv_role":"kv_producer","kv_connector_extra_config":{
+           "tilert_model":"glm5_2","tilert_max_seq_len":1048576,"tilert_transport":"mooncake"}}
+  decode:
+    engine:
+      type: tilert
+      served_model_name: glm5_2
+    container: ghcr.io/tile-ai/tilert-rocm-decode:0.1.6
+    nodes: 1
+    workers: 1
+    gpus: 8
+    env:
+      TILERT_ROLE: decode
+      TILERT_WEIGHTS_DIR: /models/GLM-5.3-tilert-tp8
+      TILERT_CONVERT_LOCK_WAIT: "21600"
+      GLM5_AR_N: "2"
+    args:
+      model: glm5_2
+      model-weights-dir: /models/GLM-5.3-tilert-tp8
+      max-seq-len: 1048576
+      kv-cache-dtype: bf16
+      transport: mooncake
+      with-mtp: true
+      num-mtp: 3
+frontend:
+  type: tilert-router
+  container_image: ghcr.io/tile-ai/tilert-rocm-decode:0.1.6
+  enable_multiple_frontends: false
+  env:
+    TILERT_ROLE: router
+  args:
+    parser: none
+    model-path: /model
+    queue-timeout: 1800
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+srun_options:
+  mem: "0"
+  container-writable: ""
+  container-remap-root: ""
+health_check:
+  max_attempts: 2160
+  interval_seconds: 5
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    CONC: "1"
+    SERVED_MODEL_NAME: glm5_2
+    RESULT_DIR: /infmax-workspace/LOGS/agentic
+    AGENTIC_OUTPUT_DIR: /infmax-workspace
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    HF_HOME: /logs/hf_cache
+    TRANSFORMERS_VERBOSITY: error
+    TOKENIZERS_PARALLELISM: "false"
diff --git a/inferencex-e2e/configs/amd-master.yaml b/inferencex-e2e/configs/amd-master.yaml
index 74c709949e..71cadbc3af 100644
--- a/inferencex-e2e/configs/amd-master.yaml
+++ b/inferencex-e2e/configs/amd-master.yaml
@@ -1406,21 +1406,8 @@ dsv41flash-fp4-mi355x-vllm-agentic-dspark:
       - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
       - { tp: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/vllm/mi355x-fp4-mtp/agentic.yaml }
 
-# Speculative decoding on an agentic scenario must run with simulated
-# synthetic acceptance at the committed golden AL for this model, thinking mode
-# and draft length (docs/PR_REVIEW_CHECKLIST.md), and a submission may not
-# substitute its own target. Nothing about that target is written here:
-# AGENTS.md forbids hard-coding an acceptance length in a master config, so
-# server_tilert.sh reads infx/golden_al_distribution/glm5.3_mtp.yaml at launch and
-# fails the run if the curve or the draft length is missing.
-# The curve is consumed in the same units as every other framework here, and
-# AgentX replays run with thinking on.
-#
-# NOTE FOR REVIEWERS: that curve is GLM-5.2's, copied because no SPEED-Bench run
-# on GLM-5.3 exists and the two share a base. It is committed as provisional and
-# labelled as such in the file. Flagging it rather than letting it read as a
-# measured 5.3 curve -- please say if you would rather see a measured curve, a
-# non-MTP agentic entry, or a waiver.
+# The golden acceptance curve remains the provisional GLM-5.2-derived curve
+# documented in infx/golden_al_distribution/glm5.3_mtp.yaml.
 glm5.3-fp8-mi355x-tilert-agentic:
   image: ghcr.io/tile-ai/tilert-rocm-decode:0.1.6
   model: zai-org/GLM-5.3
@@ -1443,16 +1430,12 @@ glm5.3-fp8-mi355x-tilert-agentic:
           ep: 1
           dp-attn: false
           additional-settings:
-          - "PREFILL_IMAGE=ghcr.io/tile-ai/tilert-rocm-prefill:0.1.6"
-          - "PREFILL_NODES=1"
+          - "CONFIG_FILE=recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
           ep: 1
           dp-attn: false
-          additional-settings:
-          - "DECODE_NODES=1"
-          - "TILERT_EXTRA_ENV=GLM5_AR_N=2"
 
 # User-selected official DeepSeek V4.1 MI35X model preview, pinned by digest.
 # Official MI350X TP4/EP4 candidate; compare against the separate TP4/EP1 vLLM baseline.
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index a0ecd2fc71..4ea1b57c50 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9003,3 +9003,9 @@
   description:
     - "Update B300 vLLM AgentX to DSpark6 on a new image with a sampled concurrency grid and per-mode --kv-cache-memory-bytes pins."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3477
+
+- config-keys:
+    - glm5.3-fp8-mi355x-tilert-agentic
+  description:
+    - "Port MI355X GLM-5.3 TileRT AgentX to native srt-slurm per-role engines."
+  pr-link: TBD
diff --git a/inferencex-e2e/runners/launch_mi355x-amds.sh b/inferencex-e2e/runners/launch_mi355x-amds.sh
index 7f313f8fb9..4e3bf2222e 100644
--- a/inferencex-e2e/runners/launch_mi355x-amds.sh
+++ b/inferencex-e2e/runners/launch_mi355x-amds.sh
@@ -56,11 +56,15 @@ if [[ "$EXECUTION_PATH" == multinode && -n "${CONFIG_FILE:-}" ]]; then
     # Reuse a provisioned image when one exists; otherwise Pyxis imports it.
     SQUASH_FILE="$SRT_SHARED_BASE/containers/$(printf '%s' "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
     [[ -f "$SQUASH_FILE" ]] || SQUASH_FILE="$IMAGE"
+    SRT_CLUSTER_ARGS=()
+    if [[ "$FRAMEWORK" == tilert ]]; then
+        SRT_CLUSTER_ARGS+=(--model GLM-5.3 /it-share/data/GLM-5.3 --mount /it-share/data /models)
+    fi
     SLURM_ACCOUNT="$USER" SLURM_PARTITION=compute NGINX_SQUASH_FILE=nginx:1.27.4 \
         write_srt_cluster_config mi355x-amds srtslurm.yaml 0 \
         --var SRT_DEFAULT_TIME_LIMIT 01:00:00 --var GITHUB_WORKSPACE "$GITHUB_WORKSPACE" \
         --container "$IMAGE" "$SQUASH_FILE" \
-        --mount /it-share/aiperf-cache /aiperf_mmap_cache || exit 1
+        --mount /it-share/aiperf-cache /aiperf_mmap_cache "${SRT_CLUSTER_ARGS[@]}" || exit 1
     make setup ARCH=x86_64
     export INFMAX_WORKSPACE="$GITHUB_WORKSPACE"
 

From e67168459a2073e1c3ba19f8d51050e7f9c66eb0 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 15:27:34 -0500
Subject: [PATCH 2/7] chore: link port validation PR

---
 inferencex-e2e/perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 4ea1b57c50..694509bf90 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9008,4 +9008,4 @@
     - glm5.3-fp8-mi355x-tilert-agentic
   description:
     - "Port MI355X GLM-5.3 TileRT AgentX to native srt-slurm per-role engines."
-  pr-link: TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3552

From d8c13305757103d141409801abff05a2591e4773 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 15:44:08 -0500
Subject: [PATCH 3/7] fix(tilert): reuse prepared AMD weights without a write
 lock

---
 .../configs/glm5.3-tilert-rocm.sh                 | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
index 1c636f8229..fe58737801 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
@@ -48,13 +48,16 @@ sys.exit(0 if int(metadata.get("num_mtp", 0)) >= 1 else 1)
 PY
 }
 
-mkdir -p "$TILERT_WEIGHTS_DIR"
-exec 9>"$TILERT_WEIGHTS_DIR/.convert.lock"
-flock -w "$TILERT_CONVERT_LOCK_WAIT" 9
 if ! weights_ready; then
-    python3 -m tilert.models.glm_5_2_rocm.weight_converter \
-        --model_dir /model --save_dir "$TILERT_WEIGHTS_DIR" --device cuda:7 --num_mtp 3
-    weights_ready
+    mkdir -p "$TILERT_WEIGHTS_DIR"
+    exec 9>"$TILERT_WEIGHTS_DIR/.convert.lock"
+    flock -w "$TILERT_CONVERT_LOCK_WAIT" 9
+    if ! weights_ready; then
+        python3 -m tilert.models.glm_5_2_rocm.weight_converter \
+            --model_dir /model --save_dir "$TILERT_WEIGHTS_DIR" --device cuda:7 --num_mtp 3
+        weights_ready
+    fi
+    exec 9>&-
 fi
 
 for file in /model/*; do

From c2f90607a33d4a5f2483c3a5b6feffe5a2f8ab06 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 15:52:10 -0500
Subject: [PATCH 4/7] refactor(tilert): keep AMD setup limited to installation

---
 .../configs/glm5.3-tilert-rocm.sh             | 43 -------------------
 .../agentx/disagg-1p1d-tp8-mtp.yaml           |  2 -
 inferencex-e2e/runners/launch_mi355x-amds.sh  |  4 ++
 3 files changed, 4 insertions(+), 45 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
index fe58737801..e9906ebb2c 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/glm5.3-tilert-rocm.sh
@@ -26,46 +26,3 @@ if [[ "$TILERT_ROLE" != prefill ]]; then
         python3 -m pip install --quiet --no-cache-dir 'transformers>=4.56'
     fi
 fi
-
-if [[ "$TILERT_ROLE" != decode ]]; then
-    exit 0
-fi
-check_env_vars TILERT_WEIGHTS_DIR TILERT_CONVERT_LOCK_WAIT
-
-weights_ready() {
-    local rank
-    for rank in {0..7}; do
-        [[ -f "$TILERT_WEIGHTS_DIR/rank$rank/model.safetensors.index.json" ]] || return 1
-    done
-    [[ -f "$TILERT_WEIGHTS_DIR/tilert_meta.json" ]] || return 1
-    python3 - "$TILERT_WEIGHTS_DIR/tilert_meta.json" <<'PY'
-import json
-import sys
-
-with open(sys.argv[1]) as source:
-    metadata = json.load(source)
-sys.exit(0 if int(metadata.get("num_mtp", 0)) >= 1 else 1)
-PY
-}
-
-if ! weights_ready; then
-    mkdir -p "$TILERT_WEIGHTS_DIR"
-    exec 9>"$TILERT_WEIGHTS_DIR/.convert.lock"
-    flock -w "$TILERT_CONVERT_LOCK_WAIT" 9
-    if ! weights_ready; then
-        python3 -m tilert.models.glm_5_2_rocm.weight_converter \
-            --model_dir /model --save_dir "$TILERT_WEIGHTS_DIR" --device cuda:7 --num_mtp 3
-        weights_ready
-    fi
-    exec 9>&-
-fi
-
-for file in /model/*; do
-    [[ -f "$file" ]] || continue
-    name="${file##*/}"
-    [[ "$name" == *.safetensors || "$name" == model.safetensors.index.json ]] && continue
-    [[ -e "$TILERT_WEIGHTS_DIR/$name" ]] || cp -p "$file" "$TILERT_WEIGHTS_DIR/$name"
-done
-[[ -f "$TILERT_WEIGHTS_DIR/chat_template.jinja" ]]
-[[ -f "$TILERT_WEIGHTS_DIR/tokenizer_config.json" || -f "$TILERT_WEIGHTS_DIR/tokenizer.json" ]]
-echo "TileRT TP8/MTP weights ready: $TILERT_WEIGHTS_DIR"
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
index b9215a3759..68b8a81220 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
@@ -57,8 +57,6 @@ roles:
     gpus: 8
     env:
       TILERT_ROLE: decode
-      TILERT_WEIGHTS_DIR: /models/GLM-5.3-tilert-tp8
-      TILERT_CONVERT_LOCK_WAIT: "21600"
       GLM5_AR_N: "2"
     args:
       model: glm5_2
diff --git a/inferencex-e2e/runners/launch_mi355x-amds.sh b/inferencex-e2e/runners/launch_mi355x-amds.sh
index 4e3bf2222e..efd0383db0 100644
--- a/inferencex-e2e/runners/launch_mi355x-amds.sh
+++ b/inferencex-e2e/runners/launch_mi355x-amds.sh
@@ -58,6 +58,10 @@ if [[ "$EXECUTION_PATH" == multinode && -n "${CONFIG_FILE:-}" ]]; then
     [[ -f "$SQUASH_FILE" ]] || SQUASH_FILE="$IMAGE"
     SRT_CLUSTER_ARGS=()
     if [[ "$FRAMEWORK" == tilert ]]; then
+        if [[ ! -r /it-share/data/GLM-5.3-tilert-tp8/tilert_meta.json ]]; then
+            echo "Missing prepared TileRT checkpoint: /it-share/data/GLM-5.3-tilert-tp8" >&2
+            exit 1
+        fi
         SRT_CLUSTER_ARGS+=(--model GLM-5.3 /it-share/data/GLM-5.3 --mount /it-share/data /models)
     fi
     SLURM_ACCOUNT="$USER" SLURM_PARTITION=compute NGINX_SQUASH_FILE=nginx:1.27.4 \

From 15b341df57e4629ee0773b56f96660532fa9ca3f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 18:31:39 -0500
Subject: [PATCH 5/7] fix(tilert): allow AgentX client to fetch its dataset

---
 .../glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml     | 1 +
 1 file changed, 1 insertion(+)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
index 68b8a81220..5c30a63054 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
@@ -97,5 +97,6 @@ benchmark:
     AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     HF_HOME: /logs/hf_cache
+    HF_HUB_OFFLINE: "0"
     TRANSFORMERS_VERBOSITY: error
     TOKENIZERS_PARALLELISM: "false"

From 34217c72ebc1869129dbb23bcd99598438065454 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 20:35:16 -0500
Subject: [PATCH 6/7] fix(tilert): clear inherited offline mode for AgentX

---
 .../glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml    | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
index 5c30a63054..4882a1c910 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
@@ -97,6 +97,6 @@ benchmark:
     AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     HF_HOME: /logs/hf_cache
-    HF_HUB_OFFLINE: "0"
+    HF_HUB_OFFLINE: ""
     TRANSFORMERS_VERBOSITY: error
     TOKENIZERS_PARALLELISM: "false"

From cdb425edba274924da2cc5c4b147c51628625f1d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 28 Sep 2026 21:28:38 -0500
Subject: [PATCH 7/7] fix(tilert): preserve AgentX tokenizer and context
 selection

---
 .../glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml    | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
index 4882a1c910..066183afa7 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.3/tilert/mi355x-fp8/agentx/disagg-1p1d-tp8-mtp.yaml
@@ -91,7 +91,9 @@ benchmark:
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
   env:
     CONC: "1"
+    MODEL: /model
     SERVED_MODEL_NAME: glm5_2
+    AIPERF_MAX_CONTEXT_LENGTH: "1048576"
     RESULT_DIR: /infmax-workspace/LOGS/agentic
     AGENTIC_OUTPUT_DIR: /infmax-workspace
     AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache