diff --git a/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-1p1d-tp8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-1p1d-tp8-8k1k.yaml
new file mode 100644
index 000000000..f3cd36eb0
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-1p1d-tp8-8k1k.yaml
@@ -0,0 +1,136 @@
+# DeepSeek-V4-Pro 1P1D ATOM P/D serving on MI355X. Native AToMesh
+# routes prefill/decode traffic while ATOM transfers KV state through
+# Mooncake's default RDMA transport. The custom benchmark preserves the
+# legacy 8k/1k search and request-length distribution.
+
+name: "mi355x-atom-dsv4-disagg-1p1d-tp8-8k1k"
+
+model:
+  path: "hf:deepseek-ai/DeepSeek-V4-Pro"
+  container: "rocm/atom-dev:nightly_202609031453"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "rocm/atom-dev:nightly_202609031453"
+  frameworks:
+    atomesh: "f95ef3ec30a16876a676e0af96d60f8e0605c7c5"
+
+slurm:
+  time_limit: "08:00:00"
+
+resources:
+  gpu_type: "mi355x"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: atomesh
+  enable_multiple_frontends: false
+  args:
+    policy: random
+    log-level: info
+    disable-health-check: true
+    disable-circuit-breaker: true
+    prometheus-port: 29100
+
+environment: &runtime_environment
+  HF_HOME: "/hf_hub_cache"
+  HF_HUB_CACHE: "/hf_hub_cache/hub"
+  HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+  # Use the official image's ROCm/Mooncake library paths (Python 3.12).
+  PYTHONUNBUFFERED: "1"
+  PYTHONDONTWRITEBYTECODE: "1"
+  SAFETENSORS_FAST_GPU: "1"
+  VLLM_LOG_LEVEL: "WARNING"
+  ATOM_LOG_LEVEL: "WARNING"
+  AITER_LOG_LEVEL: "WARNING"
+  LOG_LEVEL: "WARNING"
+  LOGLEVEL: "WARNING"
+
+backend:
+  type: atom
+  connector: mooncake
+  prefill_environment: &worker_environment
+    ATOM_MOE_GU_ITLV: "1"
+    AITER_BF16_FP8_MOE_BOUND: "0"
+  decode_environment: *worker_environment
+  atom_config:
+    prefill:
+      # Upstream workflow MAX_MODEL_LEN = ISL + OSL + 256.
+      max-model-len: 9472
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.85
+      max-num-seqs: 256
+      block-size: 16
+      no-enable_prefix_caching: true
+      trust-remote-code: true
+    decode:
+      max-model-len: 9472
+      cudagraph-capture-sizes: "[1,2,4,8,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,136,144,152,160,168,176,184,192,200,208,216,224,232,240,248,256]"
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.85
+      max-num-seqs: 128
+      block-size: 16
+      no-enable_prefix_caching: true
+      trust-remote-code: true
+
+# The legacy exclusive allocation exposed the full 128-core, unlimited-memory
+# node; do not inherit the shared small-model profile's 32-core / 128G limits.
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+  exclusive: ""
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    archive_runtime_logs() {
+      tar -C /logs -czf "${result_root}/runtime-logs.tar.gz" . 2>/dev/null || true
+    }
+    trap archive_runtime_logs EXIT
+    for concurrency in ${CONC_LIST}; do
+      num_prompts=$((concurrency * 10))
+      if (( num_prompts < 16 )); then num_prompts=16; fi
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model deepseek-ai/DeepSeek-V4-Pro \
+        --tokenizer deepseek-ai/DeepSeek-V4-Pro \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio "${RANDOM_RANGE_RATIO:-0.8}" \
+        --num-warmups "$((concurrency * 2))" \
+        --num-prompts "${num_prompts}" \
+        --max-concurrency "${concurrency}" \
+        --request-rate inf \
+        --ignore-eos \
+        --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result \
+        --result-dir "${result_root}/fixed-seq" \
+        --result-filename "dsv4-atom-mi355x-disagg-1p1d-tp8-isl8192-osl1024-c${concurrency}.json"
+    done
+  env:
+    <<: *runtime_environment
diff --git a/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-2p1d-dpa-tp8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-2p1d-dpa-tp8-8k1k.yaml
new file mode 100644
index 000000000..22c7ad775
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-2p1d-dpa-tp8-8k1k.yaml
@@ -0,0 +1,140 @@
+# DeepSeek-V4-Pro 2P1D ATOM P/D serving on MI355X. Each logical endpoint
+# occupies one TP8 node; the two prefill endpoints and decode endpoint use
+# data-parallel attention/TBO as in the legacy production configuration.
+
+name: "mi355x-atom-dsv4-disagg-2p1d-dpa-tp8-8k1k"
+
+model:
+  path: "hf:deepseek-ai/DeepSeek-V4-Pro"
+  container: "rocm/atom-dev:nightly_202609031453"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "rocm/atom-dev:nightly_202609031453"
+  frameworks:
+    atomesh: "f95ef3ec30a16876a676e0af96d60f8e0605c7c5"
+
+slurm:
+  time_limit: "08:00:00"
+
+resources:
+  gpu_type: "mi355x"
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+frontend:
+  type: atomesh
+  enable_multiple_frontends: false
+  args:
+    policy: random
+    log-level: info
+    disable-health-check: true
+    disable-circuit-breaker: true
+    prometheus-port: 29100
+
+environment: &runtime_environment
+  HF_HOME: "/hf_hub_cache"
+  HF_HUB_CACHE: "/hf_hub_cache/hub"
+  HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+  # Use the official image's ROCm/Mooncake library paths (Python 3.12).
+  PYTHONUNBUFFERED: "1"
+  PYTHONDONTWRITEBYTECODE: "1"
+  SAFETENSORS_FAST_GPU: "1"
+  VLLM_LOG_LEVEL: "WARNING"
+  ATOM_LOG_LEVEL: "WARNING"
+  AITER_LOG_LEVEL: "WARNING"
+  LOG_LEVEL: "WARNING"
+  LOGLEVEL: "WARNING"
+
+backend:
+  type: atom
+  connector: mooncake
+  prefill_environment: &worker_environment
+    ATOM_MOE_GU_ITLV: "1"
+    AITER_BF16_FP8_MOE_BOUND: "0"
+    GPU_MAX_HW_QUEUES: "5"
+    ATOM_CPU_AFFINITY: "1"
+  decode_environment: *worker_environment
+  atom_config:
+    prefill:
+      # Upstream workflow MAX_MODEL_LEN = ISL + OSL + 256.
+      max-model-len: 9472
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.85
+      max-num-seqs: 256
+      block-size: 16
+      no-enable_prefix_caching: true
+      enable-dp-attention: true
+      enable-tbo: true
+      trust-remote-code: true
+    decode:
+      max-model-len: 9472
+      cudagraph-capture-sizes: "[1,2,4,8,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,136,144,152,160,168,176,184,192,200,208,216,224,232,240,248,256]"
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.85
+      max-num-seqs: 2048
+      block-size: 16
+      no-enable_prefix_caching: true
+      enable-dp-attention: true
+      enable-tbo: true
+      trust-remote-code: true
+
+# The legacy exclusive allocation exposed the full 128-core, unlimited-memory
+# node; do not inherit the shared small-model profile's 32-core / 128G limits.
+sbatch_directives:
+  cpus-per-task: "128"
+  mem: "0"
+  exclusive: ""
+
+srun_options:
+  container-writable: ""
+  container-remap-root: ""
+  mem: "0"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 5
+
+benchmark:
+  type: custom
+  command: |
+    set -euo pipefail
+    result_root="/results/${SLURM_JOB_ID}"
+    mkdir -p "${result_root}/fixed-seq"
+    archive_runtime_logs() {
+      tar -C /logs -czf "${result_root}/runtime-logs.tar.gz" . 2>/dev/null || true
+    }
+    trap archive_runtime_logs EXIT
+    for concurrency in ${CONC_LIST}; do
+      num_prompts=$((concurrency * 10))
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model deepseek-ai/DeepSeek-V4-Pro \
+        --tokenizer deepseek-ai/DeepSeek-V4-Pro \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio "${RANDOM_RANGE_RATIO:-0.8}" \
+        --num-warmups "$((concurrency * 2))" \
+        --num-prompts "${num_prompts}" \
+        --max-concurrency "${concurrency}" \
+        --request-rate inf \
+        --ignore-eos \
+        --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result \
+        --result-dir "${result_root}/fixed-seq" \
+        --result-filename "dsv4-atom-mi355x-disagg-2p1d-dpa-tp8-isl8192-osl1024-c${concurrency}.json"
+    done
+  env:
+    <<: *runtime_environment
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0d5dbbec0..3937a8f00 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -796,13 +796,13 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
 # MXFP8 runs from TP=4 on gfx950; block size 128 is mandatory for MSA.
 dsv4-fp4-mi355x-atom-disagg:
-  image: rocm/atom-dev:nightly_202606101403
+  image: rocm/atom-dev:nightly_202609031453
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
-  runner: mi355x
+  runner: cluster:mi355x-amds
   precision: fp4
   framework: atom-disagg
-  router: { name: atomesh, version: "087b82d9c1f630e79149ba37e6213257ec9a76f8" }
+  router: { name: atomesh, version: "f95ef3ec30a16876a676e0af96d60f8e0605c7c5" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -820,6 +820,7 @@ dsv4-fp4-mi355x-atom-disagg:
           dp-attn: true
           additional-settings:
           - "PREFILL_NODES=2"
+          - "CONFIG_FILE=recipes/atom/deepseek-v4/mi355x/disagg-2p1d-dpa-tp8-8k1k.yaml"
         decode:
           num-worker: 1
           tp: 8
@@ -836,6 +837,7 @@ dsv4-fp4-mi355x-atom-disagg:
           dp-attn: false
           additional-settings:
           - "PREFILL_NODES=1"
+          - "CONFIG_FILE=recipes/atom/deepseek-v4/mi355x/disagg-1p1d-tp8-8k1k.yaml"
         decode:
           num-worker: 1
           tp: 8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a0f56d8a7..756c5f761 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6885,3 +6885,23 @@
     - "Remove the mandatory host-repair hook, RDMA/GPU-drain helper scripts, and separate image/model staging allocation; use provisioned caches and native Pyxis container lifecycle."
     - "Keep InferenceX-specific benchmark/eval metadata and artifact collection in a portable adapter while preserving the existing serving and workload settings."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2542
+
+- config-keys:
+    - dsv4-fp4-mi355x-atom-disagg
+  description:
+    - "Port the DeepSeek-V4-Pro FP4 MI355X ATOM disaggregated 8k/1k submission from the configuration-specific AMD launcher to srt-slurm ATOM orchestration and native AToMesh routing."
+    - "Preserve all original search points: 1P1D TP8 at concurrency 4, 8, 16, 32, 64, and 128; 2P1D DPA+TP8 at concurrency 256, 512, 768, 1024, and 2048."
+    - "Use the unchanged InferenceX benchmark_serving.py through the custom benchmark contract and remove the exact legacy wrapper plus its obsolete deprecated 1k/1k entry."
+    - "Keep the original image, Mooncake default RDMA transport, random length ratio 0.8, decode graph capture sizes, router options, and TP8 / DP-attention / TBO settings."
+    - "Use the ATOM image's bundled native AToMesh binary directly, without Infera source overlays or Infera routing, and archive the complete runtime logs from the custom benchmark."
+    - "Keep rocm/atom-dev:nightly_202606101403 and the original AToMesh revision 087b82d9c1f630e79149ba37e6213257ec9a76f8; use the shared launcher and srt-slurm PR #7 from base PR #2542."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2627
+
+- config-keys:
+    - dsv4-fp4-mi355x-atom-disagg
+  description:
+    - "Use the unmodified official rocm/atom-dev:nightly_202609031453 image and its native AToMesh revision f95ef3ec30a16876a676e0af96d60f8e0605c7c5; no source patches, overlays, or custom image."
+    - "The June image ignored max_completion_tokens after AToMesh normalized max_tokens, causing the 2P1D GSM8K decode worker to exceed its context capacity. The upstream ATOM API fix is included in this nightly."
+    - "Use the image-provided ROCm/Mooncake library paths instead of the retired Python 3.10 path; keep all explicit engine options, topologies, concurrency points, benchmark inputs, and eval thresholds unchanged."
+    - "Validate the shared single-entrypoint launcher and native Pyxis lifecycle without host-repair, RDMA-preflight, or GPU-drain scripts."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2627