diff --git a/benchmarks/multi_node/dsv4_fp4_mi355x_atom-disagg.sh b/benchmarks/multi_node/dsv4_fp4_mi355x_atom-disagg.sh
deleted file mode 100644
index d17d1a323..000000000
--- a/benchmarks/multi_node/dsv4_fp4_mi355x_atom-disagg.sh
+++ /dev/null
@@ -1,83 +0,0 @@
-#!/usr/bin/env bash
-
-source "$(dirname "$0")/../benchmark_lib.sh"
-
-check_env_vars \
-    CONC_LIST \
-    ISL \
-    OSL \
-    IMAGE \
-    SPEC_DECODING \
-    MODEL_PATH \
-    PREFILL_NUM_WORKERS \
-    PREFILL_TP \
-    PREFILL_EP \
-    PREFILL_DP_ATTN \
-    DECODE_NUM_WORKERS \
-    DECODE_TP \
-    DECODE_EP \
-    DECODE_DP_ATTN \
-    PREFILL_NODES \
-    DECODE_NODES \
-    RANDOM_RANGE_RATIO \
-    FRAMEWORK
-
-if [[ -n "$SLURM_JOB_ID" ]]; then
-  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
-fi
-
-set -x
-
-# Use upstreamed multi_node scripts (no external clone needed)
-cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
-
-# Set up SGL launch script-specific environment variables
-export TIME_LIMIT="08:00:00"
-export MODEL_PATH=$MODEL_PATH
-export MODEL_NAME=$MODEL_NAME
-export CONTAINER_IMAGE=$IMAGE
-
-if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
-export PREFILL_ENABLE_EP=false
-else
-export PREFILL_ENABLE_EP=true
-fi
-
-if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
-export PREFILL_ENABLE_DP=true
-else
-export PREFILL_ENABLE_DP=false
-fi
-
-if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
-export DECODE_ENABLE_EP=false
-else
-export DECODE_ENABLE_EP=true
-fi
-
-if [[ "$DECODE_DP_ATTN" == "true" ]]; then
-export DECODE_ENABLE_DP=true
-else
-export DECODE_ENABLE_DP=false
-fi
-
-# Launch jobs based on ISL/OSL
-# Replace ' ' in CONC_LIST with 'x' such that the concurrency list is represented
-# by a list of numbers delimited by 'x'. This is because of how the underlying launch script
-# expects the concurrencies.
-JOB_ID=$(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/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..6d100edc2
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-1p1d-tp8-8k1k.yaml
@@ -0,0 +1,124 @@
+# DeepSeek-V4-Pro 1P1D ATOM P/D serving on MI355X. Infera provides
+# discovery and KV-aware routing; ATOM transfers the KV state through
+# Mooncake. The custom benchmark preserves the legacy 8k/1k search space.
+
+name: "mi355x-atom-dsv4-disagg-1p1d-tp8-8k1k"
+
+model:
+  path: "hf:deepseek-ai/DeepSeek-V4-Pro"
+  container: "infera-atom-v0.1.1"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "rocm/infera:atom-v0.1.1"
+  frameworks:
+    atom: "0.1.4.dev113+g5837907f3"
+    infera: "0.0.0"
+
+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: infera
+  enable_multiple_frontends: false
+  env:
+    PYTHONPATH: "/atom-source:/infera-source"
+  args:
+    router-policy: kv-aware
+
+backend:
+  type: atom
+  connector: mooncake
+  # The production-green legacy sweep used Mooncake RDMA on rdma0-rdma7.
+  # The pinned ATOM source contains the ROCm local-device filtering fix and
+  # selects the per-GPU RDMA device; srt-slurm supplies the cluster GID index.
+  mooncake_protocol: rdma
+  enable_kv_events: true
+  prefill_environment: &worker_environment
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache/hub"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+    PYTHONPATH: "/atom-source:/infera-source"
+    PYTHONUNBUFFERED: "1"
+    OMP_NUM_THREADS: "1"
+    ATOM_MOE_GU_ITLV: "1"
+    AITER_BF16_FP8_MOE_BOUND: "0"
+  decode_environment: *worker_environment
+  atom_config:
+    prefill:
+      kv_cache_dtype: fp8
+      # Leave registration/runtime headroom after the initial 0.85 run reached
+      # server registration with less than 6 MiB free and failed a 6 MiB calloc.
+      gpu-memory-utilization: 0.84
+      max-num-seqs: 256
+      block-size: 16
+      no-enable_prefix_caching: true
+      trust-remote-code: true
+    decode:
+      kv_cache_dtype: fp8
+      gpu-memory-utilization: 0.84
+      max-num-seqs: 128
+      block-size: 16
+      no-enable_prefix_caching: true
+      trust-remote-code: true
+
+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 "${result_root}/runtime-logs.tar.gz" . 2>/dev/null || true
+    }
+    trap archive_runtime_logs EXIT
+    for concurrency in 4 8 16 32 64 128; do
+      num_prompts=$((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 1.0 \
+        --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:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache/hub
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache/hub
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..676dae777
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/atom/deepseek-v4/mi355x/disagg-2p1d-dpa-tp8-8k1k.yaml
@@ -0,0 +1,127 @@
+# 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: "infera-atom-v0.1.1"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "rocm/infera:atom-v0.1.1"
+  frameworks:
+    atom: "0.1.4.dev113+g5837907f3"
+    infera: "0.0.0"
+
+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: infera
+  enable_multiple_frontends: false
+  env:
+    PYTHONPATH: "/atom-source:/infera-source"
+  args:
+    router-policy: kv-aware
+
+backend:
+  type: atom
+  connector: mooncake
+  # Match the production-green legacy sweep: this high-throughput search uses
+  # Mooncake RDMA across rdma0-rdma7. TCP saturates during c256 and causes the
+  # decode endpoint to miss transfer deadlines even with connection pooling.
+  mooncake_protocol: rdma
+  enable_kv_events: true
+  prefill_environment: &worker_environment
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache/hub"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache/hub"
+    PYTHONPATH: "/atom-source:/infera-source"
+    PYTHONUNBUFFERED: "1"
+    OMP_NUM_THREADS: "1"
+    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:
+      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:
+      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
+
+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 256 512 768 1024 2048; do
+      num_prompts=$((concurrency * 10))
+      python3 /infmax-workspace/utils/bench_serving/benchmark_serving.py \
+        --backend openai \
+        --base-url "http://${SRT_FRONTEND_HOST}:${SRT_FRONTEND_PORT}" \
+        --endpoint /v1/completions \
+        --model deepseek-ai/DeepSeek-V4-Pro \
+        --tokenizer deepseek-ai/DeepSeek-V4-Pro \
+        --dataset-name random \
+        --random-input-len 8192 \
+        --random-output-len 1024 \
+        --random-range-ratio 1.0 \
+        --num-warmups "$((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:
+    HF_HOME: /hf_hub_cache
+    HF_HUB_CACHE: /hf_hub_cache/hub
+    HUGGINGFACE_HUB_CACHE: /hf_hub_cache/hub
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 15ec613e4..a9ea0162a 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -1381,17 +1381,14 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=1"
 
-# MiniMax-M3 MXFP8 MI355X recipe:
-# 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/infera:atom-v0.1.1
   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: infera, version: "0.1.1" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -1400,8 +1397,11 @@ dsv4-fp4-mi355x-atom-disagg:
     - isl: 8192
       osl: 1024
       search-space:
-      # 2P1D DPA+TP8 
-      - conc-list: [ 256, 512, 768, 1024, 2048 ]
+      # The custom srt-slurm recipes own the exact concurrency loops. Keeping
+      # one matrix row per topology avoids launching the complete loop once per
+      # concurrency while preserving every original measurement point.
+      # 2P1D DPA+TP8
+      - conc-list: [ 1 ]
         prefill:
           num-worker: 2
           tp: 8
@@ -1409,6 +1409,7 @@ 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
@@ -1416,8 +1417,8 @@ dsv4-fp4-mi355x-atom-disagg:
           dp-attn: true
           additional-settings:
           - "DECODE_NODES=1"
-      # 1P1D TP8 
-      - conc-list: [ 4, 8, 16, 32, 64, 128 ]
+      # 1P1D TP8
+      - conc-list: [ 1 ]
         prefill:
           num-worker: 1
           tp: 8
@@ -1425,6 +1426,7 @@ 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
@@ -1432,7 +1434,9 @@ dsv4-fp4-mi355x-atom-disagg:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
-      # 1P1D TP8
+# MiniMax-M3 MXFP8 MI300X recipe:
+# https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
+# MXFP8 runs from TP=4 on gfx942; block size 128 is mandatory for MSA.
 minimaxm3-fp8-mi300x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/configs/deprecated/amd-1k1k-master.yaml b/configs/deprecated/amd-1k1k-master.yaml
index 3457698c3..d573e91be 100644
--- a/configs/deprecated/amd-1k1k-master.yaml
+++ b/configs/deprecated/amd-1k1k-master.yaml
@@ -1331,38 +1331,6 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
           - "DECODE_NODES=1"
           - "DECODE_MTP_SIZE=1"
 
-dsv4-fp4-mi355x-atom-disagg:
-  image: rocm/atom-dev:nightly_202606101403
-  model: deepseek-ai/DeepSeek-V4-Pro
-  model-prefix: dsv4
-  runner: mi355x
-  precision: fp4
-  framework: atom-disagg
-  router: { name: atomesh, version: "087b82d9c1f630e79149ba37e6213257ec9a76f8" }
-  kv-p2p-transfer: mooncake
-  multinode: true
-  disagg: true
-  scenarios:
-    fixed-seq-len:
-    - isl: 1024
-      osl: 1024
-      search-space:
-      - conc-list: [ 4, 8, 16, 32, 64, 128, 256, 512, 1024 ]
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "PREFILL_NODES=1"
-        decode:
-          num-worker: 1
-          tp: 8
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "DECODE_NODES=1"
-
 # MiniMax-M3 MXFP8 MI355X recipe:
 # https://github.com/vllm-project/recipes/commit/2a3728ed9892debfd767a72a58ebc90b33f186e5
 # MXFP8 runs from TP=4 on gfx950; block size 128 is mandatory for MSA.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 635670c08..cb72db136 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5807,3 +5807,14 @@
     - "Overlay the exact upstream Infera KV-event decoder fix from AMD-AGI/Infera#114 so ATOM's tagged-map events populate the router cache view during pre-release hardware validation."
     - "Serialize shared Infera and ATOM source staging per checkout and publish completed clones atomically so concurrent aggregate and disaggregated validation jobs cannot corrupt one another."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2558
+
+- 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 Infera KV-aware 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."
+    - "Restore the production-proven Mooncake RDMA transport used by the all-green legacy 8k/1k sweep; the TCP bring-up path saturated at c256 and lost the decode worker despite connection pooling."
+    - "Lower only the 1P1D prefill/decode gpu-memory-utilization from 0.85 to 0.84 after both roles registered their Mooncake chunks with less than 6 MiB free and failed a 6 MiB runtime allocation; keep the already-healthy 2P1D DPA path at 0.85."
+    - "Relay KV events from every ATOM data-parallel rank through one worker-local Infera endpoint, close the custom benchmark concurrency loops, and use a shell function for reliable runtime-log archival."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2627
diff --git a/runners/launch_mi355x-amds-srt.sh b/runners/launch_mi355x-amds-srt.sh
index 72ab623ae..2d541eddc 100755
--- a/runners/launch_mi355x-amds-srt.sh
+++ b/runners/launch_mi355x-amds-srt.sh
@@ -7,7 +7,7 @@ set -euo pipefail
 SRT_SLURM_REPOSITORY="https://github.com/SemiAnalysisAI/srt-slurm.git"
 SRT_SLURM_COMMIT="5ecfb13d1ba0960045482f1ef006312d8729d37a"
 INFERA_REPOSITORY="https://github.com/cquil11/Infera.git"
-INFERA_COMMIT="8ed8f1728c745d4e91ba9eaa09ed81159aa57e41"
+INFERA_COMMIT="503dcbd1b60229a5a508fd3da45994c56a2b40ec"
 ATOM_REPOSITORY="https://github.com/cquil11/ATOM.git"
 ATOM_COMMIT="2ab42bc2c64d1ad04f698c396da48473e71a6dbb"
 SLURM_PARTITION="compute"
diff --git a/utils/test_mi355x_atom_srt_contract.py b/utils/test_mi355x_atom_srt_contract.py
index 50a9e7f44..b4495f4af 100644
--- a/utils/test_mi355x_atom_srt_contract.py
+++ b/utils/test_mi355x_atom_srt_contract.py
@@ -75,7 +75,7 @@ def test_matrix_rows_route_only_through_the_atom_srt_launcher():
         "CONFIG_FILE=recipes/atom/qwen3-0.6b/mi355x/disagg-1p1d-fixed-seq.yaml"
     ]
     assert "5ecfb13d1ba0960045482f1ef006312d8729d37a" in launcher
-    assert "8ed8f1728c745d4e91ba9eaa09ed81159aa57e41" in launcher
+    assert "503dcbd1b60229a5a508fd3da45994c56a2b40ec" in launcher
     assert "2ab42bc2c64d1ad04f698c396da48473e71a6dbb" in launcher
     assert "ensure_git_checkout()" in launcher
     assert 'exec {lock_fd}>"\\${target}.lock"' in launcher