diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 8b9b38f2e5..e8ffd55a55 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -3297,7 +3297,7 @@ build_replay_cmd() {
     # necessarily a valid HF repo id (e.g. "Qwen3.5-397B-A17B-NVFP4-V2" vs
     # the real "nvidia/Qwen3.5-397B-A17B-NVFP4-V2"), which 404s tokenizer
     # loading. Always pass the real HF id explicitly.
-    REPLAY_CMD+=" --tokenizer $MODEL"
+    REPLAY_CMD+=" --tokenizer ${AIPERF_TOKENIZER:-$MODEL}"
     REPLAY_CMD+=" --concurrency $CONC"
     REPLAY_CMD+=" --benchmark-duration $duration"
     REPLAY_CMD+=" --stats-interval 30"
diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_b200_sglang.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_b200_sglang.sh
new file mode 100644
index 0000000000..fa16ad4fed
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_b200_sglang.sh
@@ -0,0 +1,2 @@
+#!/usr/bin/env bash
+exec bash "$(dirname "$0")/qwen3.5_fp8_blackwell_powerx.sh" "$@"
diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_b300_sglang.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_b300_sglang.sh
new file mode 100644
index 0000000000..fa16ad4fed
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_b300_sglang.sh
@@ -0,0 +1,2 @@
+#!/usr/bin/env bash
+exec bash "$(dirname "$0")/qwen3.5_fp8_blackwell_powerx.sh" "$@"
diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_blackwell_powerx.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_blackwell_powerx.sh
new file mode 100644
index 0000000000..09fb3958ec
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_blackwell_powerx.sh
@@ -0,0 +1,151 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Controlled PowerX AgentX arm shared by B200 and B300: TP4, FP8, no MTP,
+# with GPU prefix reuse and no CPU KV offload.
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars TP CONC EP_SIZE KV_OFFLOADING RESULT_DIR DURATION
+if [[ "$TP" != 4 || "$EP_SIZE" != 1 || "${PP_SIZE:-1}" != 1 ||
+      "${DCP_SIZE:-1}" != 1 || "${PCP_SIZE:-1}" != 1 ||
+      "${DP_ATTENTION:-false}" != false || "${SPEC_DECODING:-none}" != none ]]; then
+    echo "Error: the PowerX arm requires TP4/EP1, no context/pipeline/attention parallelism, and no MTP" >&2
+    exit 1
+fi
+require_agentic_kv_offload_none
+
+export MODEL="Qwen/Qwen3.5-397B-A17B-FP8"
+export MODEL_REVISION="ea5b4f81096f3901c91dea97f81324302495781d"
+export EVAL_FRAMEWORK="lm-eval"
+export EVAL_TASKS_DIR="$INFERENCEX_REPO_ROOT/utils/evals/gsm8k.yaml"
+export ENABLE_AGENTX_POWER=1
+export REQUIRE_POWER=1
+export AIPERF_FAILED_REQUEST_THRESHOLD=0
+export AIPERF_LIVE_FAILED_REQUEST_THRESHOLD=0
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="sglang:"
+export AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES=0
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126_256k
+for simulation_var in ${!SGLANG_SIMULATE_ACC_@}; do
+    unset "$simulation_var"
+done
+
+mkdir -p "$RESULT_DIR"
+SERVER_LOG="$RESULT_DIR/server.log"
+install_agentic_deps
+# Staged MODEL_PATH directories do not prove a checkpoint revision. Reuse a
+# pinned HF cache snapshot when available, otherwise download that revision.
+"$AIPERF_HF_CLI" download "$MODEL" --revision "$MODEL_REVISION" --dry-run \
+    | tee "$RESULT_DIR/powerx_model_cache.txt"
+MODEL_PATH=$("$AIPERF_HF_CLI" download "$MODEL" --revision "$MODEL_REVISION")
+export MODEL_PATH
+export AIPERF_TOKENIZER="$MODEL_PATH"
+resolve_trace_source
+
+verify_trace_revision() {
+    "$AIPERF_PYTHON" - <<'PY' | tee -a "$RESULT_DIR/powerx_dataset_revision.txt"
+from huggingface_hub import HfApi
+
+revision = HfApi().dataset_info("semianalysisai/cc-traces-weka-062126-256k").sha
+print(revision)
+if revision != "8fecd2fc56694469f758f0afbbb6335ad3043740":
+    raise SystemExit("PowerX AgentX dataset revision changed")
+PY
+}
+verify_trace_revision
+
+{
+    printf 'model=%s\nmodel_revision=%s\nmodel_path=%s\nimage=%s\n' \
+        "$MODEL" "$MODEL_REVISION" "$MODEL_PATH" "${IMAGE:-unknown}"
+    printf 'inferencex_commit=%s\naiperf_commit=%s\n' \
+        "$(git -c safe.directory="$INFERENCEX_REPO_ROOT" -C "$INFERENCEX_REPO_ROOT" rev-parse HEAD)" \
+        "$(git -c safe.directory="$AIPERF_DIR" -C "$AIPERF_DIR" rev-parse HEAD)"
+    printf 'tp=%s\nep=%s\nconcurrency=%s\nduration=%s\nagentx_fast=%s\ncuda_visible_devices=%s\n' \
+        "$TP" "$EP_SIZE" "$CONC" "$DURATION" "${AIPERF_EXPERIMENTAL_FAST:-0}" "${CUDA_VISIBLE_DEVICES:-unset}"
+    sha256sum "$MODEL_PATH/config.json" "$MODEL_PATH/tokenizer_config.json"
+    nvidia-smi --query-gpu=index,uuid,name,driver_version,power.limit --format=csv
+} > "$RESULT_DIR/powerx_runtime.txt"
+python3 -m pip freeze > "$RESULT_DIR/powerx_server_packages.txt"
+"$AIPERF_UV_BIN" pip freeze --python "$AIPERF_PYTHON" > "$RESULT_DIR/powerx_client_packages.txt"
+
+# Concurrency counts session trees; retain room for subagent requests.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS="$CONC"
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+
+export TORCH_CUDA_ARCH_LIST="10.0"
+export PYTHONNOUSERSITE=1
+export NCCL_NVLS_ENABLE=1
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+export SGLANG_TIMEOUT_KEEP_ALIVE=1800
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --tp "$TP"
+    --dp 1
+    --ep-size "$EP_SIZE"
+    --enable-symm-mem
+    --quantization fp8
+    --kv-cache-dtype fp8_e4m3
+    --mamba-ssm-dtype bfloat16
+    --attention-backend trtllm_mha
+    --moe-runner-backend flashinfer_trtllm
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --mem-fraction-static 0.80
+    --stream-interval 50
+    --scheduler-recv-interval 10
+    --tokenizer-worker-num 6
+    --tokenizer-path "$MODEL_PATH"
+    --reasoning-parser qwen3
+    --tool-call-parser qwen3_coder
+    --enable-metrics
+    --enable-cache-report
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    capture_cache_metrics
+    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+capture_cache_metrics() {
+    {
+        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
+        curl -fsS "http://localhost:$PORT/metrics" 2>/dev/null \
+            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|token_usage|num_requests_running|num_requests_waiting))' \
+            || true
+    } >> "$SERVER_LOG"
+}
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+capture_cache_metrics
+
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --apply-chat-template"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+    verify_trace_revision
+fi
diff --git a/benchmarks/single_node/agentic/qwen3.5_fp8_mi355x_sglang.sh b/benchmarks/single_node/agentic/qwen3.5_fp8_mi355x_sglang.sh
new file mode 100755
index 0000000000..1549e48106
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp8_mi355x_sglang.sh
@@ -0,0 +1,132 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Controlled PowerX AgentX replay: FP8 TP4, native decoding, HBM prefix reuse.
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC EP_SIZE KV_OFFLOADING RESULT_DIR DURATION
+require_agentic_kv_offload_none
+if [[ "$TP" != 4 || "$EP_SIZE" != 1 || "${SPEC_DECODING:-none}" != none ]]; then
+    echo "Error: PowerX requires TP4/EP1 without speculative decoding" >&2
+    exit 1
+fi
+
+export EVAL_FRAMEWORK=lm-eval
+export EVAL_TASKS_DIR="$INFERENCEX_REPO_ROOT/utils/evals/gsm8k.yaml"
+export ENABLE_AGENTX_POWER=1
+export REQUIRE_POWER=1
+export AIPERF_FAILED_REQUEST_THRESHOLD=0
+export AIPERF_LIVE_FAILED_REQUEST_THRESHOLD=0
+export AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES=0
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="sglang:"
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126_256k
+unset SGLANG_SIMULATE_ACC_LEN SGLANG_SIMULATE_ACC_METHOD SGLANG_SIMULATE_ACC_TOKEN_MODE
+
+mkdir -p "$RESULT_DIR"
+install_agentic_deps
+guard_powerx_dataset_revision() {
+    "$AIPERF_PYTHON" - <<'PY'
+from huggingface_hub import HfApi
+
+dataset = "semianalysisai/cc-traces-weka-062126-256k"
+expected = "8fecd2fc56694469f758f0afbbb6335ad3043740"
+actual = HfApi().dataset_info(dataset).sha
+if actual != expected:
+    raise SystemExit(f"PowerX dataset revision changed: {actual} != {expected}")
+print(f"{dataset}@{actual}")
+PY
+}
+guard_powerx_dataset_revision > "$RESULT_DIR/powerx_dataset_revision.txt"
+resolve_trace_source
+export MODEL_REVISION=ea5b4f81096f3901c91dea97f81324302495781d
+"$AIPERF_HF_CLI" download "$MODEL" --revision "$MODEL_REVISION" --dry-run \
+    | tee "$RESULT_DIR/powerx_model_cache.txt"
+MODEL_PATH=$("$AIPERF_HF_CLI" download "$MODEL" --revision "$MODEL_REVISION")
+export MODEL_PATH
+export AIPERF_TOKENIZER="$MODEL_PATH"
+
+SERVER_LOG="$RESULT_DIR/server.log"
+{
+    printf 'model=%s\nmodel_revision=%s\nmodel_path=%s\nimage=%s\n' \
+        "$MODEL" "$MODEL_REVISION" "$MODEL_PATH" "${IMAGE:-unknown}"
+    printf 'ROCR_VISIBLE_DEVICES=%s\nHIP_VISIBLE_DEVICES=%s\nCUDA_VISIBLE_DEVICES=%s\n' \
+        "${ROCR_VISIBLE_DEVICES:-}" "${HIP_VISIBLE_DEVICES:-}" "${CUDA_VISIBLE_DEVICES:-}"
+    git -c safe.directory="$INFERENCEX_REPO_ROOT" -C "$INFERENCEX_REPO_ROOT" rev-parse HEAD
+    git -c safe.directory="$AIPERF_DIR" -C "$AIPERF_DIR" rev-parse HEAD
+    sha256sum "$MODEL_PATH/config.json" "$MODEL_PATH/tokenizer_config.json"
+} > "$RESULT_DIR/powerx_runtime.txt"
+amd-smi static --json > "$RESULT_DIR/powerx_gpu_identity.json"
+python3 -m pip freeze > "$RESULT_DIR/powerx_server_packages.txt"
+"$AIPERF_UV_BIN" pip freeze --python "$AIPERF_PYTHON" > "$RESULT_DIR/powerx_client_packages.txt"
+
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "SGLang server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS="$CONC"
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+
+export PYTHONNOUSERSITE=1
+export SGLANG_USE_AITER=1
+export SGLANG_USE_AITER_UNIFIED_ATTN=1
+export SGLANG_MAMBA_SSM_DTYPE=bfloat16
+export SGLANG_TIMEOUT_KEEP_ALIVE=1800
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --tp "$TP"
+    --dp 1
+    --ep-size "$EP_SIZE"
+    --attention-backend aiter
+    --enable-aiter-allreduce-fusion
+    --quantization fp8
+    --kv-cache-dtype fp8_e4m3
+    --mamba-ssm-dtype bfloat16
+    --mem-fraction-static 0.80
+    --model-loader-extra-config '{"enable_multithread_load": true}'
+    --watchdog-timeout 1200
+    --page-size 16
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --scheduler-recv-interval 10
+    --stream-interval 50
+    --tokenizer-worker-num 6
+    --tokenizer-path "$MODEL_PATH"
+    --reasoning-parser qwen3
+    --tool-call-parser qwen3_coder
+    --enable-metrics
+    --enable-cache-report
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [ "${EVAL_ONLY:-false}" = true ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --apply-chat-template"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+    guard_powerx_dataset_revision >> "$RESULT_DIR/powerx_dataset_revision.txt"
+fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0269ddfa05..68d7672423 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -200,6 +200,19 @@ qwen3.5-fp8-mi355x-sglang:
       search-space:
       - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
 
+qwen3.5-fp8-mi355x-sglang-agentic-powerx:
+  image: lmsysorg/sglang:v0.5.16-rocm720-mi35x@sha256:54ac680bad1832b8acd469533ae66f608b525cec3449bbd5f3d0238351e9b965
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:mi355x-amds
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - { tp: 4, ep: 1, spec-decoding: none, kv-offloading: none, conc-list: [1, 4, 12] }
+
 qwen3.5-fp8-mi355x-sglang-mtp:
   image: lmsysorg/sglang-rocm:v0.5.18-rocm720-mi35x-20260828
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 0a6c86df6d..10d6c2bdd0 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1215,6 +1215,20 @@ qwen3.5-fp8-b200-sglang:
       - { tp: 8, conc-start: 4, conc-end: 4 }
       - { tp: 4, ep: 1, conc-start: 4, conc-end: 256 }
 
+# Controlled PowerX arm; AgentX matrix generation defaults to 3600 seconds.
+qwen3.5-fp8-b200-sglang-agentic-powerx:
+  image: lmsysorg/sglang:v0.5.16-cu130@sha256:7b6a35df9839fd593a94a1eaee82d7777f472225d9f3ad1f8a2e0cb2bd1785d0
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:b200-nscale
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - { tp: 4, ep: 1, spec-decoding: none, kv-offloading: none, conc-list: [1, 4, 12] }
+
 qwen3.5-fp8-b200-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: Qwen/Qwen3.5-397B-A17B-FP8
@@ -7200,6 +7214,20 @@ qwen3.5-fp4-gb300-dynamo-sglang-agentic-disagg:
           dp-attn: false
 
     # ---------- 1k1k high-throughput (wide-EP decode, EAGLE MTP) ----------
+# Controlled PowerX arm; shares the B200 launcher and 3600-second default.
+qwen3.5-fp8-b300-sglang-agentic-powerx:
+  image: lmsysorg/sglang:v0.5.16-cu130@sha256:7b6a35df9839fd593a94a1eaee82d7777f472225d9f3ad1f8a2e0cb2bd1785d0
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:b300-dsxe
+  precision: fp8
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - { tp: 4, ep: 1, spec-decoding: none, kv-offloading: none, conc-list: [1, 4, 12] }
+
 qwen3.5-fp8-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/docs/eval-agentx-procedures.md b/docs/eval-agentx-procedures.md
index 469bd4aafd..8fb71f7b76 100644
--- a/docs/eval-agentx-procedures.md
+++ b/docs/eval-agentx-procedures.md
@@ -173,6 +173,8 @@ Retain `meta_env.json`, `results*.json`, and `sample*.jsonl`. Agentic SWE-bench
 
 ## 7. Run AgentX: fast feedback versus canonical evidence
 
+The controlled PowerX keys `qwen3.5-fp8-{b200,b300,mi355x}-sglang-agentic-powerx` use TP4, native decoding, prefix reuse and no CPU offload. Qualify their serving and four-GPU power boundaries before collecting the configured one-hour runs. A recipe can set `AIPERF_TOKENIZER` to its pinned model snapshot so replay and serving use the same tokenizer; otherwise replay continues to use `MODEL`.
+
 AgentX is AIPerf `inferencex-agentx-mvp` trace replay, not a fixed-token synthetic benchmark. The checked-in default uses ten additional warmup requests per trajectory lane and the recipe's configured profile duration. `agentx-fast` forces one warmup request per lane and a 1,200-second profile. It affects single- and multi-node AgentX throughput only. Fixed-sequence throughput and evals remain canonical. Fast runs are not eligible for artifact reuse ([workflow policy](../.github/workflows/README.md#agentx-fast-mode), [fast replay settings](../benchmarks/benchmark_lib.sh#L2104-L2128)).
 
 For multi-node srt-slurm jobs, the benchmark client may run on a different host from the frontend. `agentic_srt.sh` uses an explicit `AIPERF_SERVER_URL` when supplied, otherwise derives it from `SRT_FRONTEND_HOST` and `SRT_FRONTEND_PORT`, and falls back to `localhost:$PORT` only when no remote endpoint is available. Trace replay and inter-point drain checks must use that same resolved endpoint.
@@ -338,3 +340,5 @@ Use `scancel` or process termination only with explicit approval and a concrete
 - Every backend/frontend and metrics source is represented in live evidence.
 - Fast/smoke results are labeled diagnostic. Only the canonical candidate is used for final comparison.
 - Workflow and artifact collection conclude green before success is reported.
+
+The PowerX runners convert Docker digest pins to Enroot manifest references at import time. B200 and B300 mount persistent Hugging Face caches, and the recipes record a pinned-model download dry run before loading weights.
diff --git a/docs/eval-agentx-procedures_zh.md b/docs/eval-agentx-procedures_zh.md
index 501344bf25..0618e2d685 100644
--- a/docs/eval-agentx-procedures_zh.md
+++ b/docs/eval-agentx-procedures_zh.md
@@ -173,6 +173,8 @@ gh run download "$RUN_ID" --repo SemiAnalysisAI/InferenceX \
 
 ## 7. 运行 AgentX：快速反馈与 canonical 证据
 
+受控 PowerX 配置 `qwen3.5-fp8-{b200,b300,mi355x}-sglang-agentic-powerx` 使用 TP4、原生解码、前缀复用，且不启用 CPU offload。在采集配置规定的一小时结果前，先验证服务与四张 GPU 的功耗统计边界。recipe 可将 `AIPERF_TOKENIZER` 设为固定版本的模型快照路径，让 replay 与服务使用同一 tokenizer；未设置时仍使用 `MODEL`。
+
 AgentX 是 AIPerf `inferencex-agentx-mvp` trace replay，不是固定 token 的合成 benchmark。仓库默认设置对每条 trajectory lane 额外执行十个 warmup 请求，并使用 recipe 配置的 profile 时长。`agentx-fast` 强制每条 lane 只运行一个 warmup 请求，并将 profile 设为 1,200 秒。它只影响单节点和多节点 AgentX 吞吐量；定长序列吞吐量与 eval 保持 canonical。Fast 运行不符合 artifact reuse 条件（[工作流策略](../.github/workflows/README.md#agentx-fast-mode)、[fast replay 设置](../benchmarks/benchmark_lib.sh#L2104-L2128)）。
 
 对于多节点 srt-slurm 作业，benchmark client 与 frontend 可能运行在不同主机上。`agentic_srt.sh` 会优先使用显式提供的 `AIPERF_SERVER_URL`；否则从 `SRT_FRONTEND_HOST` 和 `SRT_FRONTEND_PORT` 推导地址；仅在没有远端 endpoint 时回退到 `localhost:$PORT`。Trace replay 和并发点之间的 drain 检查必须使用同一个解析后的 endpoint。
@@ -338,3 +340,5 @@ gh run cancel <RUN_ID> --repo SemiAnalysisAI/InferenceX
 - 每个 backend/frontend 与 metrics source 都在实时证据中有所体现。
 - Fast/smoke 结果明确标为诊断用途；只有 canonical candidate 用于最终比较。
 - 在报告成功前，工作流与 artifact collection 均已得出 green 结论。
+
+PowerX runner 在导入时将 Docker digest 转换为 Enroot manifest 引用。B200 和 B300 挂载持久化 Hugging Face 缓存，配置在加载权重前记录固定模型版本的下载预检查。
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 14434c76ad..9dddd51785 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6878,3 +6878,25 @@
   description:
     - "Refresh the Kimi-K3 GB300 AgentX configurations with updated configs."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2811
+
+- config-keys:
+    - qwen3.5-fp8-b200-sglang-agentic-powerx
+    - qwen3.5-fp8-b300-sglang-agentic-powerx
+    - qwen3.5-fp8-mi355x-sglang-agentic-powerx
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add controlled Qwen3.5 FP8 TP4 AgentX power configurations for B200, B300, and MI355X with prefix caching and no speculative decoding or CPU offload."
+    - "Pin serving images and model weights, require valid GPU power, and allow the replay tokenizer to use the pinned model snapshot."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2827
+
+- config-keys:
+    - qwen3.5-fp8-b200-sglang-agentic-powerx
+    - qwen3.5-fp8-b300-sglang-agentic-powerx
+    - qwen3.5-fp8-mi355x-sglang-agentic-powerx
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Fix digest-pinned image imports for the PowerX qualification runners by reusing the Enroot manifest-reference converter."
+    - "Mount persistent HF caches on the NVIDIA single-node runners and record pinned model cache requirements before download."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2827
diff --git a/runners/container_utils.sh b/runners/container_utils.sh
new file mode 100644
index 0000000000..a1e6f209c8
--- /dev/null
+++ b/runners/container_utils.sh
@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+
+# Enroot 3.x does not parse Docker's tag@digest syntax. For digest-pinned
+# images, use its explicit registry syntax and pass the digest as the
+# manifest reference so the import remains immutable.
+enroot_uri_for_image() {
+    local image="$1"
+    local image_without_digest="$image"
+    local digest=""
+    local first_component registry repository repository_dir repository_name
+
+    if [[ "$image" == *@sha256:* ]]; then
+        image_without_digest="${image%@*}"
+        digest="${image##*@}"
+    fi
+
+    first_component="${image_without_digest%%/*}"
+    if [[ "$image_without_digest" == */* && ( "$first_component" == *.* || "$first_component" == *:* || "$first_component" == "localhost" ) ]]; then
+        registry="$first_component"
+        repository="${image_without_digest#*/}"
+    else
+        registry="registry-1.docker.io"
+        repository="$image_without_digest"
+    fi
+
+    if [[ -z "$digest" ]]; then
+        if [[ "$registry" == "registry-1.docker.io" ]]; then
+            printf 'docker://%s\n' "$image"
+        else
+            printf 'docker://%s#%s\n' "$registry" "$repository"
+        fi
+        return
+    fi
+
+    repository_dir="${repository%/*}"
+    repository_name="${repository##*/}"
+    repository_name="${repository_name%%:*}"
+    if [[ "$repository" == */* ]]; then
+        repository="${repository_dir}/${repository_name}"
+    else
+        repository="$repository_name"
+    fi
+    if [[ "$registry" == "registry-1.docker.io" && "$repository" != */* ]]; then
+        repository="library/$repository"
+    fi
+
+    printf 'docker://%s#%s:%s\n' "$registry" "$repository" "$digest"
+}
diff --git a/runners/launch_b200-nscale-compat.sh b/runners/launch_b200-nscale-compat.sh
index 15b4013aac..d93bed41ce 100644
--- a/runners/launch_b200-nscale-compat.sh
+++ b/runners/launch_b200-nscale-compat.sh
@@ -1,5 +1,7 @@
 #!/usr/bin/bash
 
+source "$(dirname "${BASH_SOURCE[0]}")/container_utils.sh"
+
 # Compatibility launcher for B200 Nscale configurations that have not yet
 # moved to the native srt-slurm path in launch_b200-nscale-slurm.sh.
 SLURM_PARTITION="${SLURM_PARTITION:-batch_1}"
@@ -570,11 +572,17 @@ else
     salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
+    # Keep pinned model snapshots and datasets across single-node jobs.
+    HF_HUB_CACHE_HOST_PATH="/data/home/sa-shared/gharunners/hf-hub-cache"
+    export HF_HUB_CACHE="${HF_HUB_CACHE:-/mnt/hf_hub_cache}"
+    mkdir -p "$HF_HUB_CACHE_HOST_PATH"
+
     # Point the bench script at the resolved MODEL_PATH instead of
     # pulling from the HF hub cache. Bench scripts skip `hf download` when
     # MODEL is a local path.
     export MODEL="$MODEL_PATH"
 
+    ENROOT_IMAGE_URI=$(enroot_uri_for_image "$IMAGE") || exit 1
     # Use flock to serialize concurrent imports to the same squash file
     # Override ENROOT_CACHE_PATH to avoid permission issues with system-wide cache on worker nodes
     srun --jobid=$JOB_ID bash -c "
@@ -586,13 +594,13 @@ else
             echo 'Squash file already exists and is valid, skipping import'
         else
             rm -f \"$SQUASH_FILE\"
-            enroot import -o \"$SQUASH_FILE\" docker://$IMAGE
+            enroot import -o \"$SQUASH_FILE\" \"$ENROOT_IMAGE_URI\"
         fi
     "
 
     srun --jobid=$JOB_ID \
         --container-image=$SQUASH_FILE \
-        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache \
+        --container-mounts=$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR,$MODEL_PATH:$MODEL_PATH,$HF_HUB_CACHE_HOST_PATH:$HF_HUB_CACHE,$AIPERF_MMAP_CACHE_HOST_PATH:/aiperf_mmap_cache \
         --no-container-mount-home \
         --container-workdir=$CONTAINER_MOUNT_DIR \
         --no-container-entrypoint --export=ALL,PORT=8888,AIPERF_DATASET_MMAP_CACHE_DIR=/aiperf_mmap_cache \
diff --git a/runners/launch_b300-dsxe.sh b/runners/launch_b300-dsxe.sh
index c229146bfa..3ba6f960dc 100755
--- a/runners/launch_b300-dsxe.sh
+++ b/runners/launch_b300-dsxe.sh
@@ -1,5 +1,7 @@
 #!/usr/bin/bash
 
+source "$(dirname "${BASH_SOURCE[0]}")/container_utils.sh"
+
 # Launcher for the B300 DSXE Slurm cluster (dsxe-sa-b300-prd0), runners run as sa-gha-runner.
 #
 # Every cluster-specific fact lives in this block. The rest of the file is generic:
@@ -87,6 +89,8 @@ import_squash_image() {
     local image_ref="$1"
     local sqsh="$2"
     local lock="${2}.lock"
+    local enroot_uri
+    enroot_uri=$(enroot_uri_for_image "$image_ref") || exit 1
 
     if unsquashfs -l "$sqsh" > /dev/null 2>&1; then
         echo "Squash file already present, skipping import: $sqsh"
@@ -102,7 +106,7 @@ import_squash_image() {
             exit 0
         fi
         rm -f \"$sqsh\"
-        enroot import -o \"$sqsh\" \"docker://$image_ref\"
+        enroot import -o \"$sqsh\" \"$enroot_uri\"
         unsquashfs -l \"$sqsh\" > /dev/null
     " || { echo "Error: enroot import failed for $image_ref -> $sqsh" >&2; exit 1; }
 
@@ -432,9 +436,10 @@ done
 find . -name '.nfs*' -delete 2>/dev/null || true
 
 else
-    # HF_HUB_CACHE is set to help with dataset download inside the container
-    # for eval jobs.
-    export HF_HUB_CACHE="$HOME/.cache/huggingface"
+    # Persist pinned snapshots and datasets even without the home mount.
+    HF_HUB_CACHE_HOST_PATH="/data/home/sa-gha-runner/hf-hub-cache"
+    export HF_HUB_CACHE="${HF_HUB_CACHE:-/mnt/hf_hub_cache}"
+    mkdir -p "$HF_HUB_CACHE_HOST_PATH"
 
     # MODEL stays the HF id for the client; MODEL_PATH is where the server reads
     # weights. Only the root holding MODEL_PATH is mounted -- mounting both roots
@@ -499,6 +504,7 @@ else
     CONTAINER_MOUNTS=(
         "$GITHUB_WORKSPACE:$CONTAINER_MOUNT_DIR"
         "$MODEL_MOUNT_DIR:$MODEL_MOUNT_DIR"
+        "$HF_HUB_CACHE_HOST_PATH:$HF_HUB_CACHE"
     )
     CONTAINER_MOUNTS_ARG=$(IFS=,; printf '%s' "${CONTAINER_MOUNTS[*]}")
 
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 9799aa3b71..48a4049675 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -16,52 +16,7 @@ SQUASH_DIR="/mnt/lustre01/users-public/sa-shared"
 POWER_SRT_SLURM_URL="https://github.com/edwingao28/srt-slurm.git"
 POWER_SRT_SLURM_PIN="6fc1bed01a0b82dae0088a105c03ce0cfb353443"
 
-# Enroot 3.x does not parse Docker's tag@digest syntax. For digest-pinned
-# images, use its explicit registry syntax and pass the digest as the
-# manifest reference so the import remains immutable.
-enroot_uri_for_image() {
-    local image="$1"
-    local image_without_digest="$image"
-    local digest=""
-    local first_component registry repository repository_dir repository_name
-
-    if [[ "$image" == *@sha256:* ]]; then
-        image_without_digest="${image%@*}"
-        digest="${image##*@}"
-    fi
-
-    first_component="${image_without_digest%%/*}"
-    if [[ "$image_without_digest" == */* && ( "$first_component" == *.* || "$first_component" == *:* || "$first_component" == "localhost" ) ]]; then
-        registry="$first_component"
-        repository="${image_without_digest#*/}"
-    else
-        registry="registry-1.docker.io"
-        repository="$image_without_digest"
-    fi
-
-    if [[ -z "$digest" ]]; then
-        if [[ "$registry" == "registry-1.docker.io" ]]; then
-            printf 'docker://%s\n' "$image"
-        else
-            printf 'docker://%s#%s\n' "$registry" "$repository"
-        fi
-        return
-    fi
-
-    repository_dir="${repository%/*}"
-    repository_name="${repository##*/}"
-    repository_name="${repository_name%%:*}"
-    if [[ "$repository" == */* ]]; then
-        repository="${repository_dir}/${repository_name}"
-    else
-        repository="$repository_name"
-    fi
-    if [[ "$registry" == "registry-1.docker.io" && "$repository" != */* ]]; then
-        repository="library/$repository"
-    fi
-
-    printf 'docker://%s#%s:%s\n' "$registry" "$repository" "$digest"
-}
+source "$(dirname "${BASH_SOURCE[0]}")/container_utils.sh"
 
 # Concurrent matrix jobs import to the same shared-FS squash path.
 # Serialize imports and atomically replace invalid images so readers never
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
index 49c3af0cf5..78ec6b76c4 100644
--- a/runners/launch_mi355x-amds.sh
+++ b/runners/launch_mi355x-amds.sh
@@ -1,5 +1,7 @@
 #!/usr/bin/env bash
 
+source "$(dirname "${BASH_SOURCE[0]}")/container_utils.sh"
+
 scancel_sync() {
     local jobid=$1
     local timeout=${2:-600}
@@ -271,6 +273,7 @@ else
 
     srun --jobid=$JOB_ID bash -c "docker stop \$(docker ps -a -q)"
 
+    ENROOT_IMAGE_URI=$(enroot_uri_for_image "$IMAGE") || exit 1
     # Use flock to serialize concurrent imports to the same squash file
     srun --jobid=$JOB_ID bash -c "
         exec 9>\"$LOCK_FILE\"
@@ -279,7 +282,7 @@ else
             echo 'Squash file already exists and is valid, skipping import'
         else
             rm -f \"$SQUASH_FILE\"
-            enroot import -o \"$SQUASH_FILE\" docker://$IMAGE
+            enroot import -o \"$SQUASH_FILE\" \"$ENROOT_IMAGE_URI\"
         fi
     "
 
diff --git a/utils/test_container_utils.py b/utils/test_container_utils.py
new file mode 100644
index 0000000000..b846fb403e
--- /dev/null
+++ b/utils/test_container_utils.py
@@ -0,0 +1,53 @@
+"""Check image references against the Enroot 3.x consumer grammar."""
+
+import subprocess
+from pathlib import Path
+
+import pytest
+
+
+HELPER = Path(__file__).resolve().parents[1] / "runners" / "container_utils.sh"
+DIGEST = "sha256:" + "a" * 64
+
+# Consumer grammar from NVIDIA/enroot v3.5.0 src/docker.sh:259-269.
+# `tag` is passed directly to /v2/<image>/manifests/<tag>, so sha256:<hex>
+# remains an immutable manifest reference without using Docker's @ syntax.
+PARSE_URI = r"""
+source "$1"
+uri=$(enroot_uri_for_image "$2") || exit
+reg_user="[[:alnum:]_.!~*\'()%\;:\&=+$,-@]+"
+reg_registry="[^#]+"
+reg_image="[[:lower:][:digit:]/._-]+"
+reg_tag="[[:alnum:]._:-]+"
+if [[ "${uri}" =~ ^docker://((${reg_user})@)?((${reg_registry})#)?(${reg_image})(:(${reg_tag}))?$ ]]; then
+    printf '%s\n' "${BASH_REMATCH[2]}" "${BASH_REMATCH[4]}" "${BASH_REMATCH[5]}" "${BASH_REMATCH[7]}"
+else
+    exit 1
+fi
+"""
+
+
+@pytest.mark.parametrize(
+    ("image", "registry", "repository", "reference"),
+    [
+        (f"lmsysorg/sglang:v0.5.16-cu130@{DIGEST}", "registry-1.docker.io", "lmsysorg/sglang", DIGEST),
+        (f"lmsysorg/sglang:v0.5.16-rocm720-mi35x@{DIGEST}", "registry-1.docker.io", "lmsysorg/sglang", DIGEST),
+        (f"lmsysorg/sglang@{DIGEST}", "registry-1.docker.io", "lmsysorg/sglang", DIGEST),
+        (f"ubuntu@{DIGEST}", "registry-1.docker.io", "library/ubuntu", DIGEST),
+        (f"ghcr.io/org/image:release@{DIGEST}", "ghcr.io", "org/image", DIGEST),
+        (f"localhost:5000/org/image:release@{DIGEST}", "localhost:5000", "org/image", DIGEST),
+        ("lmsysorg/sglang:v0.5.16-cu130", "", "lmsysorg/sglang", "v0.5.16-cu130"),
+        ("nvcr.io/nvidia/cuda:13.0", "nvcr.io", "nvidia/cuda", "13.0"),
+    ],
+)
+def test_enroot_reference_preserves_manifest_identity(
+    image: str, registry: str, repository: str, reference: str
+) -> None:
+    result = subprocess.run(
+        ["bash", "-c", PARSE_URI, "bash", str(HELPER), image],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr
+    assert result.stdout.splitlines() == ["", registry, repository, reference]