From b3f8ba49fc5455c8e341c8bae362d4f707cae2fd Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:20:07 -0500
Subject: [PATCH 01/17] Add H100 MiniMax-M3 disaggregated source validation

---
 .../disagg-1p1d-tp8-tp8-c8-agentic.yaml       | 128 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  27 ++++
 runners/launch_h100-dgxc-slurm.sh             |  33 ++++-
 .../test_generate_sweep_configs.py            |   4 +-
 4 files changed, 188 insertions(+), 4 deletions(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml
new file mode 100644
index 0000000000..171eacf265
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml
@@ -0,0 +1,128 @@
+name: "minimax-m3-vllm-disagg-h100-1p1d-tp8-tp8-c8-fp8-agentic"
+
+model:
+  path: "MiniMaxAI/MiniMax-M3-MXFP8"
+  container: "ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h"
+  precision: "fp8"
+
+identity:
+  model:
+    repo: "MiniMaxAI/MiniMax-M3-MXFP8"
+    revision: "c5454eb03678d8710e54a4e0fc681b9f3b4a3dba"
+  container:
+    image: "ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h"
+  frameworks:
+    dynamo: "1.4.0.dev20260730"
+
+dynamo:
+  install: true
+  version: "1.4.0.dev20260730"
+  request_plane: "nats"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+slurm:
+  time_limit: "5:00:00"
+
+resources:
+  gpu_type: "h100"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+  args:
+    dyn-chat-processor: "vllm"
+    trust-remote-code: true
+    tool-call-parser: "minimax_m3"
+    reasoning-parser: "minimax_m3"
+    enable-auto-tool-choice: true
+    router-mode: "kv"
+    router-kv-events: true
+    router-reset-states: true
+    router-temperature: "0"
+    router-session-affinity-ttl-secs: 14400
+    kv-cache-block-size: 128
+  env:
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+
+backend:
+  type: "vllm"
+  connector: null
+  kv_events_config: true
+
+  prefill_environment: &worker-environment
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    PYTHONNOUSERSITE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+
+  decode_environment: *worker-environment
+
+  vllm_config:
+    prefill: &worker-config
+      served-model-name: "MiniMaxAI/MiniMax-M3-MXFP8"
+      tensor-parallel-size: 8
+      trust-remote-code: true
+      enable-prefix-caching: true
+      kv-cache-metrics: true
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      cpu-offload-gb: 26
+      max-model-len: 1048576
+      language-model-only: true
+      kv-cache-dtype: "fp8"
+      attention-backend: "TRITON_ATTN"
+      max-num-seqs: 16
+      max-cudagraph-capture-size: 32
+      max-num-batched-tokens: 16384
+      safetensors-load-strategy: "lazy"
+      enable-prompt-tokens-details: true
+      default-chat-template-kwargs: '{"thinking_mode":"enabled"}'
+      reasoning-parser: "minimax_m3"
+      dyn-tool-call-parser: "minimax_m3"
+      dyn-reasoning-parser: "minimax_m3"
+    decode: *worker-config
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: "custom"
+  command: "bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh"
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_EXTRA_INPUTS: "thinking:true"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 481db4fcef..bbef99b0db 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7356,6 +7356,33 @@ minimaxm3-fp8-h100-vllm-agentic-mtp:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 3, 4, 5] }
       - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [6, 8] }
 
+minimaxm3-fp8-h100-dynamo-vllm-agentic-disagg:
+  image: ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: cluster:h100-dgxc
+  precision: fp8
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.4.0.dev20260730" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: none
+        kv-offloading: none
+        conc-list: [8]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml"
+        decode: { num-worker: 1, tp: 8, ep: 1, dp-attn: false }
+
 minimaxm3-fp8-h200-vllm-agentic-mtp:
   image: vllm/vllm-openai:v0.27.1
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index 50c6a8ad1f..eebbd58f68 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -4,6 +4,8 @@ set -e
 # System-specific configuration for H100 DGXC Slurm cluster
 SLURM_PARTITION="hpc-gpu-1"
 SLURM_ACCOUNT="customer"
+VLLM_AGENTIC_SRT_TAG="v1.0.50"
+VLLM_AGENTIC_SRT_PIN="e4019633c9e2bc25f38c44b81edf52bb0504d937"
 
 # Route spec-decoding=mtp configs to the _mtp benchmark script (parity with
 # the h200 launchers, which have carried SPEC_SUFFIX since #392).
@@ -33,8 +35,16 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             echo "Unsupported model prefix/precision for dynamo-trt: $MODEL_PREFIX/$PRECISION"
             exit 1
         fi
+    elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
+        if [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+            export MODEL_PATH="/mnt/nfs/sa-shared/gharunners/hf-hub-cache/models--MiniMaxAI--MiniMax-M3-MXFP8/snapshots/c5454eb03678d8710e54a4e0fc681b9f3b4a3dba"
+            export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
+        else
+            echo "Unsupported model prefix/precision for dynamo-vllm: $MODEL_PREFIX/$PRECISION"
+            exit 1
+        fi
     else
-        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang"
+        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, dynamo-vllm"
         exit 1
     fi
 
@@ -45,8 +55,19 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         rm -rf "$SRT_REPO_DIR"
     fi
 
+    if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+        git clone --branch "$VLLM_AGENTIC_SRT_TAG" --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+        test "$(git rev-parse HEAD)" = "$VLLM_AGENTIC_SRT_PIN" || {
+            echo "Error: srt-slurm $VLLM_AGENTIC_SRT_TAG resolved to an unexpected commit" >&2
+            exit 1
+        }
+        mkdir -p recipes/vllm/minimax-m3/h100-fp8/agentic
+        cp -rT \
+            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic" \
+            recipes/vllm/minimax-m3/h100-fp8/agentic
     # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" ]]; then
+    elif [[ "$IS_AGENTIC" == "1" ]]; then
         git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR"
     else
@@ -93,6 +114,13 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         # TRT-LLM container mapping - convert IMAGE to srt-slurm format (nvcr.io/ -> nvcr.io#)
         CONTAINER_KEY=$(echo "$IMAGE" | sed 's|nvcr.io/|nvcr.io#|')
         SQUASH_FILE="/mnt/nfs/sa-shared/containers/$(echo "$IMAGE" | sed 's|nvcr.io/||' | sed 's/[\/:@#]/+/g').sqsh"
+    elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
+        CONTAINER_KEY="$IMAGE"
+        SQUASH_FILE="/mnt/nfs/lustre/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+        if ! unsquashfs -l "$SQUASH_FILE" >/dev/null 2>&1; then
+            echo "Error: expected pre-imported vLLM image at $SQUASH_FILE" >&2
+            exit 1
+        fi
     fi
 
     export ISL="$ISL"
@@ -120,6 +148,7 @@ model_paths:
 containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
+  dynamo-vllm: "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
   latest: "${SQUASH_FILE}"
   "${CONTAINER_KEY}": "${SQUASH_FILE}"
diff --git a/utils/matrix_logic/test_generate_sweep_configs.py b/utils/matrix_logic/test_generate_sweep_configs.py
index 520e4105e0..ae70ccffbf 100644
--- a/utils/matrix_logic/test_generate_sweep_configs.py
+++ b/utils/matrix_logic/test_generate_sweep_configs.py
@@ -2359,9 +2359,9 @@ def test_kimi_minimax_trimmed_eval_matrix_covers_current_configs(
             json.dumps(manifest, separators=(',', ':')).encode()
         ).hexdigest()
 
-        assert len(manifest) == 63
+        assert len(manifest) == 64
         assert manifest_digest == (
-            '1630cdd6fbf77302ee3286b118710576aa090e1502b3e5ec482b7f537a3f1132'
+            '06d825174c26317f0299837a535d50352a1fd871fb655f5bc4136cf5a0f3cd8a'
         ), manifest_digest
         for row in rows:
             if isinstance(row['conc'], list):

From cf67f7d9c6bf48ec050fbe6bc70cbbe331286d41 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:20:47 -0500
Subject: [PATCH 02/17] Schedule H100 disaggregated source validation

---
 perf-changelog.yaml | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3dd2322a81..0b44fb5ee8 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6741,3 +6741,14 @@
     - "Configure model-specific tool and reasoning parsers for the Kimi-K3 and MiniMax-M3 serving paths so opt-in verifier requests produce structured OpenAI tool calls."
     - "Limit shared AgentX readiness and synthetic-acceptance rewrites to eval-only runs so default throughput recipe rendering remains unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2634
+
+- config-keys:
+    - minimaxm3-fp8-h100-dynamo-vllm-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add one H100 MiniMax-M3 MXFP8 AgentX disaggregated validation point at concurrency 8 with one TP8 prefill worker and one TP8 decode worker."
+    - "Transfer KV directly with NixlConnector so the patched vLLM image must export transferred cached prompt tokens with source=p2p."
+    - "Pin NVIDIA/srt-slurm v1.0.50 and collect the Prometheus endpoints of both logical vLLM workers through AIPerf."
+    - "Use ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h, matching the DRAM, NVMe, and tiered validation sweep in PR #2796."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 572b08768e44beac521f997ea4fcb4a1676d7c41 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:31:53 -0500
Subject: [PATCH 03/17] Allow cold H100 disaggregated container startup

---
 runners/launch_h100-dgxc-slurm.sh | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index eebbd58f68..5412b4296b 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -62,6 +62,9 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             echo "Error: srt-slurm $VLLM_AGENTIC_SRT_TAG resolved to an unexpected commit" >&2
             exit 1
         }
+        # The 18 GB validation image can take more than srt-slurm's default
+        # five-minute window to unpack on a cold H100 node.
+        sed -i 's/timeout=300/timeout=900/g' src/srtctl/cli/do_sweep.py
         mkdir -p recipes/vllm/minimax-m3/h100-fp8/agentic
         cp -rT \
             "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic" \

From 0bb3a24c210677e81ce2ea826bcf6055b2f6c8fa Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:47:23 -0500
Subject: [PATCH 04/17] Revert "Allow cold H100 disaggregated container
 startup"

This reverts commit 572b08768e44beac521f997ea4fcb4a1676d7c41.
---
 runners/launch_h100-dgxc-slurm.sh | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index 5412b4296b..eebbd58f68 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -62,9 +62,6 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             echo "Error: srt-slurm $VLLM_AGENTIC_SRT_TAG resolved to an unexpected commit" >&2
             exit 1
         }
-        # The 18 GB validation image can take more than srt-slurm's default
-        # five-minute window to unpack on a cold H100 node.
-        sed -i 's/timeout=300/timeout=900/g' src/srtctl/cli/do_sweep.py
         mkdir -p recipes/vllm/minimax-m3/h100-fp8/agentic
         cp -rT \
             "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic" \

From 1f500885e400b407672ad15a0ef5fa812ad7a067 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:47:24 -0500
Subject: [PATCH 05/17] Revert "Schedule H100 disaggregated source validation"

This reverts commit cf67f7d9c6bf48ec050fbe6bc70cbbe331286d41.
---
 perf-changelog.yaml | 11 -----------
 1 file changed, 11 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0b44fb5ee8..3dd2322a81 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6741,14 +6741,3 @@
     - "Configure model-specific tool and reasoning parsers for the Kimi-K3 and MiniMax-M3 serving paths so opt-in verifier requests produce structured OpenAI tool calls."
     - "Limit shared AgentX readiness and synthetic-acceptance rewrites to eval-only runs so default throughput recipe rendering remains unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2634
-
-- config-keys:
-    - minimaxm3-fp8-h100-dynamo-vllm-agentic-disagg
-  scenario-type:
-    - agentic-coding
-  description:
-    - "Add one H100 MiniMax-M3 MXFP8 AgentX disaggregated validation point at concurrency 8 with one TP8 prefill worker and one TP8 decode worker."
-    - "Transfer KV directly with NixlConnector so the patched vLLM image must export transferred cached prompt tokens with source=p2p."
-    - "Pin NVIDIA/srt-slurm v1.0.50 and collect the Prometheus endpoints of both logical vLLM workers through AIPerf."
-    - "Use ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h, matching the DRAM, NVMe, and tiered validation sweep in PR #2796."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 4308bee22cd76590938d1af262e106c4197ce547 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:47:25 -0500
Subject: [PATCH 06/17] Revert "Add H100 MiniMax-M3 disaggregated source
 validation"

This reverts commit b3f8ba49fc5455c8e341c8bae362d4f707cae2fd.
---
 .../disagg-1p1d-tp8-tp8-c8-agentic.yaml       | 128 ------------------
 configs/nvidia-master.yaml                    |  27 ----
 runners/launch_h100-dgxc-slurm.sh             |  33 +----
 .../test_generate_sweep_configs.py            |   4 +-
 4 files changed, 4 insertions(+), 188 deletions(-)
 delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml
deleted file mode 100644
index 171eacf265..0000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml
+++ /dev/null
@@ -1,128 +0,0 @@
-name: "minimax-m3-vllm-disagg-h100-1p1d-tp8-tp8-c8-fp8-agentic"
-
-model:
-  path: "MiniMaxAI/MiniMax-M3-MXFP8"
-  container: "ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h"
-  precision: "fp8"
-
-identity:
-  model:
-    repo: "MiniMaxAI/MiniMax-M3-MXFP8"
-    revision: "c5454eb03678d8710e54a4e0fc681b9f3b4a3dba"
-  container:
-    image: "ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h"
-  frameworks:
-    dynamo: "1.4.0.dev20260730"
-
-dynamo:
-  install: true
-  version: "1.4.0.dev20260730"
-  request_plane: "nats"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
-slurm:
-  time_limit: "5:00:00"
-
-resources:
-  gpu_type: "h100"
-  gpus_per_node: 8
-  prefill_nodes: 1
-  decode_nodes: 1
-  prefill_workers: 1
-  decode_workers: 1
-  gpus_per_prefill: 8
-  gpus_per_decode: 8
-
-infra:
-  etcd_nats_dedicated_node: false
-  nats_max_payload_mb: 32
-
-frontend:
-  type: "dynamo"
-  enable_multiple_frontends: false
-  args:
-    dyn-chat-processor: "vllm"
-    trust-remote-code: true
-    tool-call-parser: "minimax_m3"
-    reasoning-parser: "minimax_m3"
-    enable-auto-tool-choice: true
-    router-mode: "kv"
-    router-kv-events: true
-    router-reset-states: true
-    router-temperature: "0"
-    router-session-affinity-ttl-secs: 14400
-    kv-cache-block-size: 128
-  env:
-    DYN_TCP_REQUEST_TIMEOUT: "60"
-
-backend:
-  type: "vllm"
-  connector: null
-  kv_events_config: true
-
-  prefill_environment: &worker-environment
-    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
-    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
-    VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300"
-    VLLM_FLOAT32_MATMUL_PRECISION: "high"
-    VLLM_LOG_STATS_INTERVAL: "1"
-    PYTHONNOUSERSITE: "1"
-    NCCL_CUMEM_ENABLE: "1"
-    NCCL_NVLS_ENABLE: "1"
-    UCX_MEMTYPE_CACHE: "n"
-
-  decode_environment: *worker-environment
-
-  vllm_config:
-    prefill: &worker-config
-      served-model-name: "MiniMaxAI/MiniMax-M3-MXFP8"
-      tensor-parallel-size: 8
-      trust-remote-code: true
-      enable-prefix-caching: true
-      kv-cache-metrics: true
-      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
-      block-size: 128
-      gpu-memory-utilization: 0.90
-      cpu-offload-gb: 26
-      max-model-len: 1048576
-      language-model-only: true
-      kv-cache-dtype: "fp8"
-      attention-backend: "TRITON_ATTN"
-      max-num-seqs: 16
-      max-cudagraph-capture-size: 32
-      max-num-batched-tokens: 16384
-      safetensors-load-strategy: "lazy"
-      enable-prompt-tokens-details: true
-      default-chat-template-kwargs: '{"thinking_mode":"enabled"}'
-      reasoning-parser: "minimax_m3"
-      dyn-tool-call-parser: "minimax_m3"
-      dyn-reasoning-parser: "minimax_m3"
-    decode: *worker-config
-
-sbatch_directives:
-  cpus-per-task: "144"
-  mem: "0"
-
-srun_options:
-  container-remap-root: ""
-
-benchmark:
-  type: "custom"
-  command: "bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh"
-  env:
-    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
-    RESULT_DIR: "/logs/agentic"
-    PORT: "8000"
-    IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
-    AIPERF_EXTRA_INPUTS: "thinking:true"
-    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
-    AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0"
-    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
-    HF_HUB_CACHE: "/hf_hub_cache"
-    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index bbef99b0db..481db4fcef 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7356,33 +7356,6 @@ minimaxm3-fp8-h100-vllm-agentic-mtp:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 3, 4, 5] }
       - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [6, 8] }
 
-minimaxm3-fp8-h100-dynamo-vllm-agentic-disagg:
-  image: ttl.sh/cquil11-vllm-tier-d67b417bca-pr53087-20260901:24h
-  model: MiniMaxAI/MiniMax-M3-MXFP8
-  model-prefix: minimaxm3
-  runner: cluster:h100-dgxc
-  precision: fp8
-  framework: dynamo-vllm
-  router: { name: dynamo-router, version: "1.4.0.dev20260730" }
-  kv-p2p-transfer: nixl
-  multinode: true
-  disagg: true
-  scenarios:
-    agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
-      - spec-decoding: none
-        kv-offloading: none
-        conc-list: [8]
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3/h100-fp8/agentic/disagg-1p1d-tp8-tp8-c8-agentic.yaml"
-        decode: { num-worker: 1, tp: 8, ep: 1, dp-attn: false }
-
 minimaxm3-fp8-h200-vllm-agentic-mtp:
   image: vllm/vllm-openai:v0.27.1
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index eebbd58f68..50c6a8ad1f 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -4,8 +4,6 @@ set -e
 # System-specific configuration for H100 DGXC Slurm cluster
 SLURM_PARTITION="hpc-gpu-1"
 SLURM_ACCOUNT="customer"
-VLLM_AGENTIC_SRT_TAG="v1.0.50"
-VLLM_AGENTIC_SRT_PIN="e4019633c9e2bc25f38c44b81edf52bb0504d937"
 
 # Route spec-decoding=mtp configs to the _mtp benchmark script (parity with
 # the h200 launchers, which have carried SPEC_SUFFIX since #392).
@@ -35,16 +33,8 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             echo "Unsupported model prefix/precision for dynamo-trt: $MODEL_PREFIX/$PRECISION"
             exit 1
         fi
-    elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
-        if [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
-            export MODEL_PATH="/mnt/nfs/sa-shared/gharunners/hf-hub-cache/models--MiniMaxAI--MiniMax-M3-MXFP8/snapshots/c5454eb03678d8710e54a4e0fc681b9f3b4a3dba"
-            export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
-        else
-            echo "Unsupported model prefix/precision for dynamo-vllm: $MODEL_PREFIX/$PRECISION"
-            exit 1
-        fi
     else
-        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, dynamo-vllm"
+        echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang"
         exit 1
     fi
 
@@ -55,19 +45,8 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         rm -rf "$SRT_REPO_DIR"
     fi
 
-    if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
-        git clone --branch "$VLLM_AGENTIC_SRT_TAG" --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR"
-        test "$(git rev-parse HEAD)" = "$VLLM_AGENTIC_SRT_PIN" || {
-            echo "Error: srt-slurm $VLLM_AGENTIC_SRT_TAG resolved to an unexpected commit" >&2
-            exit 1
-        }
-        mkdir -p recipes/vllm/minimax-m3/h100-fp8/agentic
-        cp -rT \
-            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/h100-fp8/agentic" \
-            recipes/vllm/minimax-m3/h100-fp8/agentic
     # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    elif [[ "$IS_AGENTIC" == "1" ]]; then
+    if [[ "$IS_AGENTIC" == "1" ]]; then
         git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR"
     else
@@ -114,13 +93,6 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         # TRT-LLM container mapping - convert IMAGE to srt-slurm format (nvcr.io/ -> nvcr.io#)
         CONTAINER_KEY=$(echo "$IMAGE" | sed 's|nvcr.io/|nvcr.io#|')
         SQUASH_FILE="/mnt/nfs/sa-shared/containers/$(echo "$IMAGE" | sed 's|nvcr.io/||' | sed 's/[\/:@#]/+/g').sqsh"
-    elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
-        CONTAINER_KEY="$IMAGE"
-        SQUASH_FILE="/mnt/nfs/lustre/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
-        if ! unsquashfs -l "$SQUASH_FILE" >/dev/null 2>&1; then
-            echo "Error: expected pre-imported vLLM image at $SQUASH_FILE" >&2
-            exit 1
-        fi
     fi
 
     export ISL="$ISL"
@@ -148,7 +120,6 @@ model_paths:
 containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
-  dynamo-vllm: "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
   latest: "${SQUASH_FILE}"
   "${CONTAINER_KEY}": "${SQUASH_FILE}"
diff --git a/utils/matrix_logic/test_generate_sweep_configs.py b/utils/matrix_logic/test_generate_sweep_configs.py
index ae70ccffbf..520e4105e0 100644
--- a/utils/matrix_logic/test_generate_sweep_configs.py
+++ b/utils/matrix_logic/test_generate_sweep_configs.py
@@ -2359,9 +2359,9 @@ def test_kimi_minimax_trimmed_eval_matrix_covers_current_configs(
             json.dumps(manifest, separators=(',', ':')).encode()
         ).hexdigest()
 
-        assert len(manifest) == 64
+        assert len(manifest) == 63
         assert manifest_digest == (
-            '06d825174c26317f0299837a535d50352a1fd871fb655f5bc4136cf5a0f3cd8a'
+            '1630cdd6fbf77302ee3286b118710576aa090e1502b3e5ec482b7f537a3f1132'
         ), manifest_digest
         for row in rows:
             if isinstance(row['conc'], list):

From c221f799761e8a7cbe63725d5c3e679cc9506b14 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 10:52:12 -0500
Subject: [PATCH 07/17] Use proven GB300 disaggregated validation point

---
 configs/nvidia-master.yaml | 104 +------------------------------------
 perf-changelog.yaml        |  10 ++++
 2 files changed, 11 insertions(+), 103 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 481db4fcef..f1caf39962 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7891,7 +7891,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg:
           - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml"
 dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
-  image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f
+  image: ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
@@ -7905,40 +7905,6 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
     - search-space:
       # Keep the checked-in recipes on real MTP verification. Throughput jobs
       # inject the committed golden AL at launch; EVAL_ONLY leaves them real.
-      - spec-decoding: mtp
-        conc-list: [1152]
-        router: { name: dynamo-router, version: "1.3.0.dev20260720" }
-        prefill:
-          num-worker: 2
-          tp: 8
-          ep: 8
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-dep12-c1152-mtp-agentic.yaml"
-        decode:
-          num-worker: 1
-          tp: 12
-          ep: 12
-          dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [1024]
-        router: { name: dynamo-router, version: "1.3.0.dev20260720" }
-        prefill:
-          num-worker: 2
-          tp: 8
-          ep: 8
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-dep16-c1024-mtp-agentic.yaml"
-        decode:
-          num-worker: 1
-          tp: 16
-          ep: 16
-          dp-attn: true
       - spec-decoding: mtp
         conc-list: [256]
         router: { name: dynamo-router, version: "1.3.0.dev20260720" }
@@ -7956,74 +7922,6 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
           tp: 8
           ep: 8
           dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [512]
-        router: { name: dynamo-router, version: "1.4.0" }
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 8
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep16-c512-mtp-agentic.yaml"
-        decode:
-          num-worker: 1
-          tp: 16
-          ep: 16
-          dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [4]
-        router: { name: dynamo-router, version: "1.4.0" }
-        prefill:
-          num-worker: 1
-          tp: 4
-          ep: 4
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p4d-dep4-tp8-c4-mtp-agentic.yaml"
-        decode:
-          num-worker: 4
-          tp: 8
-          ep: 1
-          dp-attn: false
-      - spec-decoding: mtp
-        conc-list: [128]
-        router: { name: dynamo-router, version: "1.4.0" }
-        prefill:
-          num-worker: 1
-          tp: 4
-          ep: 4
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c128-mtp-agentic.yaml"
-        decode:
-          num-worker: 1
-          tp: 16
-          ep: 16
-          dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [256]
-        router: { name: dynamo-router, version: "1.4.0" }
-        prefill:
-          num-worker: 1
-          tp: 4
-          ep: 4
-          dp-attn: true
-          additional-settings:
-          - "SYNTHETIC_ACCEPTANCE=true"
-          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml"
-        decode:
-          num-worker: 1
-          tp: 16
-          ep: 16
-          dp-attn: true
 
 # GB200 mirror of the dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-{agg,disagg}
 # tuning above (PR #2571). Any worker that used 4 GPUs on GB300 uses 8 on
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3dd2322a81..c118ed19e7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6741,3 +6741,13 @@
     - "Configure model-specific tool and reasoning parsers for the Kimi-K3 and MiniMax-M3 serving paths so opt-in verifier requests produce structured OpenAI tool calls."
     - "Limit shared AgentX readiness and synthetic-acceptance rewrites to eval-only runs so default throughput recipe rendering remains unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2634
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Validate source-labeled prompt-token cache metrics on the proven GB300 vLLM disaggregated AgentX topology at concurrency 256."
+    - "Use one TP4 prefill worker and one TP8 decode worker with NIXL KV transfer; transferred prompt-token hits must export source=p2p."
+    - "Run the Python-only metric overlay built from official vLLM nightly 7c5dc571 and the HIT_PENDING fix from vLLM PR #53087."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 66e7bf46090c50ccddc3242fac7f0a98dd14003d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 11:04:17 -0500
Subject: [PATCH 08/17] Match proven GB300 c256 topology

---
 configs/nvidia-master.yaml | 8 ++++----
 perf-changelog.yaml        | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f1caf39962..8d6cc41444 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7907,7 +7907,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
       # inject the committed golden AL at launch; EVAL_ONLY leaves them real.
       - spec-decoding: mtp
         conc-list: [256]
-        router: { name: dynamo-router, version: "1.3.0.dev20260720" }
+        router: { name: dynamo-router, version: "1.4.0" }
         prefill:
           num-worker: 1
           tp: 4
@@ -7916,11 +7916,11 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
           additional-settings:
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
-          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c256-mtp-agentic.yaml"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml"
         decode:
           num-worker: 1
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: true
 
 # GB200 mirror of the dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-{agg,disagg}
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c118ed19e7..1716f3d055 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6748,6 +6748,6 @@
     - agentic-coding
   description:
     - "Validate source-labeled prompt-token cache metrics on the proven GB300 vLLM disaggregated AgentX topology at concurrency 256."
-    - "Use one TP4 prefill worker and one TP8 decode worker with NIXL KV transfer; transferred prompt-token hits must export source=p2p."
+    - "Use one TP4 prefill worker and one TP16 decode worker with NIXL KV transfer; transferred prompt-token hits must export source=p2p."
     - "Run the Python-only metric overlay built from official vLLM nightly 7c5dc571 and the HIT_PENDING fix from vLLM PR #53087."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 3057a73ebbf8382f801a443f86c71d3615977501 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 11:17:40 -0500
Subject: [PATCH 09/17] Pin validation image in GB300 recipe

---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml        | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index cd782b4a98..74310eae20 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -9,14 +9,14 @@ name: "svf-vllm-disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic"
 
 model:
   path: "deepseek-v4-pro"
-  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f"
+  container: "ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h"
   precision: "fp4"
 
 identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
-    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f"
+    image: "ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h"
   frameworks:
     dynamo: "1.4.0"
 

From cb93ced68562bbfd28374c11b8f44bdbd166eff6 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 11:34:23 -0500
Subject: [PATCH 10/17] fix: align disaggregated validation with upstream MoE
 backend

---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml   | 4 ++--
 perf-changelog.yaml                                      | 9 +++++++++
 2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index 74310eae20..eab3e1d7b7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -102,7 +102,7 @@ backend:
       tokenizer-mode: "deepseek_v4"
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
-      moe-backend: "deep_gemm_amxf4_mega_moe"
+      moe-backend: "deep_gemm_mega_moe"
       numa-bind: true
       numa-bind-nodes: [0, 0, 1, 1]
     decode:
@@ -130,7 +130,7 @@ backend:
       tokenizer-mode: "deepseek_v4"
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
-      moe-backend: "deep_gemm_amxf4_mega_moe"
+      moe-backend: "deep_gemm_mega_moe"
       numa-bind: true
       numa-bind-nodes: [0, 0, 1, 1]
   prefill_environment:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1716f3d055..d810e77915 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6751,3 +6751,12 @@
     - "Use one TP4 prefill worker and one TP16 decode worker with NIXL KV transfer; transferred prompt-token hits must export source=p2p."
     - "Run the Python-only metric overlay built from official vLLM nightly 7c5dc571 and the HIT_PENDING fix from vLLM PR #53087."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use the validation image's upstream deep_gemm_mega_moe backend for both prefill and decode workers; the older recipe's deep_gemm_amxf4_mega_moe backend name is not supported by this image."
+    - "Preserve the proven DEP4/DEP16 topology and single concurrency-256 AgentX point with its eval companion."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 7e709c0cdbf026476b75de9c860fa37710ef62fb Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 12:09:34 -0500
Subject: [PATCH 11/17] fix: collect worker metrics for custom AgentX
 validation

---
 ...b300-1p1d-dep4-dep16-c256-mtp-agentic.yaml |  1 +
 docs/eval-agentx-procedures.md                |  2 +
 docs/eval-agentx-procedures_zh.md             |  2 +
 perf-changelog.yaml                           |  9 +++
 runners/launch_gb300-nv.sh                    |  3 +
 .../patches/srt-slurm-custom-metrics.patch    | 16 +++++
 runners/test_srt_custom_metrics_patch.py      | 64 +++++++++++++++++++
 7 files changed, 97 insertions(+)
 create mode 100644 runners/patches/srt-slurm-custom-metrics.patch
 create mode 100644 runners/test_srt_custom_metrics_patch.py

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index eab3e1d7b7..2046a5fa92 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -207,6 +207,7 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:prompt_tokens_cached_by_source"
     AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     # Avoid concurrent readers observing a mismatched mmap data/index pair.
     AIPERF_DATASET_MMAP_CACHE_ENABLED: "false"
diff --git a/docs/eval-agentx-procedures.md b/docs/eval-agentx-procedures.md
index 6db5d351e8..833a411b77 100644
--- a/docs/eval-agentx-procedures.md
+++ b/docs/eval-agentx-procedures.md
@@ -286,6 +286,8 @@ Topology rules:
 - Disaggregated: inspect every prefill backend, every decode backend, and the frontend/router. A healthy decode pool does not prove prefill/KV transfer health.
 - Confirm the AIPerf command includes all `AIPERF_SERVER_METRICS_URLS`. Missing endpoints produce falsely healthy partial evidence.
 
+On GB300's pinned srt-slurm v1.0.36 custom AgentX path, set `benchmark.env.AIPERF_REQUIRED_SERVER_METRIC_PREFIX` to opt into worker endpoint discovery and require that metric prefix in the exported artifacts. The launcher applies a scoped backport; without the opt-in, the custom command otherwise discovers only the frontend.
+
 Read each endpoint directly when summaries are ambiguous:
 
 ```bash
diff --git a/docs/eval-agentx-procedures_zh.md b/docs/eval-agentx-procedures_zh.md
index 6e8c8e03c8..11c97bab6f 100644
--- a/docs/eval-agentx-procedures_zh.md
+++ b/docs/eval-agentx-procedures_zh.md
@@ -286,6 +286,8 @@ rg -n -i 'Phase |warmup|profiling|returned=|in_flight=|queue=|kv_usage=|prefix_c
 - Disaggregated：检查每个 prefill backend、每个 decode backend 以及 frontend/router。Decode pool 健康不能证明 prefill/KV transfer 健康。
 - 确认 AIPerf 命令包含所有 `AIPERF_SERVER_METRICS_URLS`；缺少 endpoint 会产生片面而虚假的健康证据。
 
+在 GB300 固定使用的 srt-slurm v1.0.36 自定义 AgentX 路径中，设置 `benchmark.env.AIPERF_REQUIRED_SERVER_METRIC_PREFIX` 可启用 worker endpoint 发现，并要求导出的 artifact 包含该指标前缀。Launcher 会应用限定范围的兼容补丁；未启用时，自定义命令默认只发现 frontend。
+
 Summary 不明确时直接读取每个 endpoint：
 
 ```bash
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d810e77915..9f40af853e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6760,3 +6760,12 @@
     - "Use the validation image's upstream deep_gemm_mega_moe backend for both prefill and decode workers; the older recipe's deep_gemm_amxf4_mega_moe backend name is not supported by this image."
     - "Preserve the proven DEP4/DEP16 topology and single concurrency-256 AgentX point with its eval companion."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Collect every prefill and decode worker's Prometheus endpoint in the custom AgentX command using srt-slurm's existing endpoint discovery."
+    - "Require the cached-token-by-source metric in exported artifacts; router-only metrics must not pass validation. Preserve the image, topology, concurrency, and eval coverage."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 91fb3b4918..ab0f04fd16 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -299,6 +299,9 @@ elif [[ "$IS_AGENTIC" == "1" ]]; then
     # Mooncake compatibility. Keep it pinned so sweeps are reproducible.
     git clone --branch v1.0.36 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
     cd "$SRT_REPO_DIR" || exit 1
+    # v1.0.36 omits worker metrics from custom AgentX commands. Opt in only
+    # for recipes requiring backend metrics, using its existing URL discovery.
+    git apply "$GITHUB_WORKSPACE/runners/patches/srt-slurm-custom-metrics.patch" || exit 1
 
     mkdir -p recipes/vllm/deepseek-v4/agentic || exit 1
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
diff --git a/runners/patches/srt-slurm-custom-metrics.patch b/runners/patches/srt-slurm-custom-metrics.patch
new file mode 100644
index 0000000000..9fb23d6c8f
--- /dev/null
+++ b/runners/patches/srt-slurm-custom-metrics.patch
@@ -0,0 +1,16 @@
+diff --git a/src/srtctl/cli/mixins/benchmark_stage.py b/src/srtctl/cli/mixins/benchmark_stage.py
+--- a/src/srtctl/cli/mixins/benchmark_stage.py
++++ b/src/srtctl/cli/mixins/benchmark_stage.py
+@@ -468,7 +468,10 @@
+-        # Add AIPerf-specific env vars for AIPerf-driven benchmarks only
+-        if isinstance(runner, AIPerfBenchmarkRunner):
++        # Custom AgentX commands opt into worker metrics through their artifact contract.
++        if isinstance(runner, AIPerfBenchmarkRunner) or (
++            runner.name == "Custom"
++            and self.config.benchmark.env.get("AIPERF_REQUIRED_SERVER_METRIC_PREFIX")
++        ):
+             env.update(self._get_aiperf_server_metrics_env())
+             if self.config.benchmark.aiperf_package:
+                 env["AIPERF_PACKAGE"] = self.config.benchmark.aiperf_package
+ 
+         return env
diff --git a/runners/test_srt_custom_metrics_patch.py b/runners/test_srt_custom_metrics_patch.py
new file mode 100644
index 0000000000..4d6d843c64
--- /dev/null
+++ b/runners/test_srt_custom_metrics_patch.py
@@ -0,0 +1,64 @@
+import runpy
+import subprocess
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+PATCH = Path(__file__).parent / "patches/srt-slurm-custom-metrics.patch"
+TARGET = "src/srtctl/cli/mixins/benchmark_stage.py"
+ORIGINAL = '''class Stage:
+    def get_env(self, runner):
+        env = {}
+        # Add AIPerf-specific env vars for AIPerf-driven benchmarks only
+        if isinstance(runner, AIPerfBenchmarkRunner):
+            env.update(self._get_aiperf_server_metrics_env())
+            if self.config.benchmark.aiperf_package:
+                env["AIPERF_PACKAGE"] = self.config.benchmark.aiperf_package
+
+        return env
+'''
+
+
+class AIPerfRunner:
+    name = "AIPerf"
+
+
+@pytest.mark.parametrize(
+    "runner,required,expected",
+    [
+        (AIPerfRunner(), "", True),
+        (SimpleNamespace(name="Custom"), "vllm:prompt_tokens_cached_by_source", True),
+        (SimpleNamespace(name="Custom"), "", False),
+        (SimpleNamespace(name="Other"), "vllm:", False),
+    ],
+)
+def test_custom_metrics_opt_in(tmp_path, runner, required, expected):
+    target = tmp_path / TARGET
+    target.parent.mkdir(parents=True)
+    target.write_text(ORIGINAL)
+    subprocess.run(["git", "apply", str(PATCH)], cwd=tmp_path, check=True)
+    symbols = runpy.run_path(
+        str(target), init_globals={"AIPerfBenchmarkRunner": AIPerfRunner}
+    )
+    stage = symbols["Stage"]()
+    stage.config = SimpleNamespace(
+        benchmark=SimpleNamespace(
+            env={"AIPERF_REQUIRED_SERVER_METRIC_PREFIX": required},
+            aiperf_package=None,
+        )
+    )
+    urls = "http://prefill:7500/metrics,http://decode:7501/metrics"
+    stage._get_aiperf_server_metrics_env = lambda: {"AIPERF_SERVER_METRICS_URLS": urls}
+    assert stage.get_env(runner) == (
+        {"AIPERF_SERVER_METRICS_URLS": urls} if expected else {}
+    )
+
+
+def test_custom_metrics_patch_rejects_unexpected_source(tmp_path):
+    target = tmp_path / TARGET
+    target.parent.mkdir(parents=True)
+    target.write_text("# Different upstream implementation\n")
+    result = subprocess.run(["git", "apply", str(PATCH)], cwd=tmp_path, capture_output=True)
+    assert result.returncode != 0
+    assert target.read_text() == "# Different upstream implementation\n"

From edc96a6d14a089c8cdcde90d105025de6bc6e8ae Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 15:35:38 -0500
Subject: [PATCH 12/17] fix(agentx): isolate P/D profiling metrics from warmup

---
 perf-changelog.yaml | 9 +++++++++
 utils/aiperf        | 2 +-
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 71fd22866e..6dac5ad99d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6801,3 +6801,12 @@
     - "Collect every prefill and decode worker's Prometheus endpoint in the custom AgentX command using srt-slurm's existing endpoint discovery."
     - "Require the cached-token-by-source metric in exported artifacts; router-only metrics must not pass validation. Preserve the image, topology, concurrency, and eval coverage."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Pin the AgentX client fix that tags automatic warmup baseline scrapes as warmup, preventing warmup counter growth from entering profiling-only token-source totals."
+    - "Preserve the proven DEP4/DEP16 topology, concurrency 256, patched vLLM image, full eval, and canonical warmup and one-hour profiling."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
diff --git a/utils/aiperf b/utils/aiperf
index 754356e9a3..dcbd942654 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 754356e9a39acc6cc6afb242d123bb57c3fb6f75
+Subproject commit dcbd942654ef87b357c7722837855dbccedc2353

From ef79a0b7765b0f12fe702edaebf64f82d3364f31 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 1 Sep 2026 15:51:03 -0500
Subject: [PATCH 13/17] fix: retain Mooncake store checkpoints in the P/D
 validation image

---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml   | 4 ++--
 configs/nvidia-master.yaml                               | 2 +-
 perf-changelog.yaml                                      | 9 +++++++++
 3 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index 2046a5fa92..eafab72c24 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -9,14 +9,14 @@ name: "svf-vllm-disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic"
 
 model:
   path: "deepseek-v4-pro"
-  container: "ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h"
+  container: "ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h"
   precision: "fp4"
 
 identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
-    image: "ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h"
+    image: "ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h"
   frameworks:
     dynamo: "1.4.0"
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index b5d6b6dbec..b905532c1e 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8049,7 +8049,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg:
           - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml"
 dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
-  image: ttl.sh/cquil11-vllm-tier-7c5dc571-pr53087-d9b6d498-arm64-20260901:24h
+  image: ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6dac5ad99d..d87e4cc64a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6810,3 +6810,12 @@
     - "Pin the AgentX client fix that tags automatic warmup baseline scrapes as warmup, preventing warmup counter growth from entering profiling-only token-source totals."
     - "Preserve the proven DEP4/DEP16 topology, concurrency 256, patched vLLM image, full eval, and canonical warmup and one-hour profiling."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Backport vLLM PR #46906 so Mooncake's external store retains sliding-window prefix checkpoints independently of HBM retention."
+    - "Use the Python-only token-source validation image b7fc8985 while preserving the proven DEP4/DEP16 topology, concurrency 256, full eval, and canonical warmup and one-hour profiling."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 0fcc29d28679f43138de4f9740f890567e1ff10f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 2 Sep 2026 20:39:41 -0500
Subject: [PATCH 14/17] fix(agentic): recover hybrid KV cache misses in P/D
 validation image

Use supporting image 4d3c4a23 with upstream hybrid load-failure recovery and the already validated async queue-progress fix. Preserve the single DEP4/DEP16 c256 point, full eval, and canonical warmup/profile. The previous run crashed on a Mooncake OBJECT_NOT_FOUND response followed by single-group tuple unpack in the scheduler. AI assistance used for diagnosis and validation.
---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml |  4 ++--
 configs/nvidia-master.yaml                             |  2 +-
 perf-changelog.yaml                                    | 10 ++++++++++
 3 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index eafab72c24..c3052eb473 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -9,14 +9,14 @@ name: "svf-vllm-disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic"
 
 model:
   path: "deepseek-v4-pro"
-  container: "ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h"
+  container: "ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h"
   precision: "fp4"
 
 identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
-    image: "ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h"
+    image: "ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h"
   frameworks:
     dynamo: "1.4.0"
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 7acfc86dca..18b01bd957 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8049,7 +8049,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg:
           - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml"
 dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
-  image: ttl.sh/cquil11-vllm-tier-pd-retention-b7fc8985-arm64-20260901:24h
+  image: ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 5b42deff27..66ef476b90 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6867,3 +6867,13 @@
     - "Backport vLLM PR #46906 so Mooncake's external store retains sliding-window prefix checkpoints independently of HBM retention."
     - "Use the Python-only token-source validation image b7fc8985 while preserving the proven DEP4/DEP16 topology, concurrency 256, full eval, and canonical warmup and one-hour profiling."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use the 4d3c4a23 Python-only validation image with upstream PR #50388 hybrid KV-load recovery, so a missing Mooncake object does not crash DeepSeek-V4's multi-group scheduler."
+    - "Keep failed loads out of cached-token source totals and include the async-load queue-progress fix already validated by the H100 offload sweep."
+    - "Preserve the existing DEP4/DEP16 topology, single concurrency-256 point, real model eval, canonical warmup, and one-hour profiling."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From dab55b2c772368bc0d537fceff3a8443a5b8b61c Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 2 Sep 2026 20:41:07 -0500
Subject: [PATCH 15/17] fix(agentic): recompute missing prefill cache loads

Set the prefill MultiConnector failure policy to recompute so optional Mooncake cache eviction races recover through the tested hybrid path. Preserve fail-closed NIXL decode transfers. AI assistance used for diagnosis and validation.
---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml   | 2 +-
 perf-changelog.yaml                                      | 9 +++++++++
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index c3052eb473..a03716c7a3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -79,7 +79,7 @@ backend:
       enable_offload: false
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
       served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
       safetensors-load-strategy: "prefetch"
       kv-cache-dtype: "fp8"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 66ef476b90..6a66d98599 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6877,3 +6877,12 @@
     - "Keep failed loads out of cached-token source totals and include the async-load queue-progress fix already validated by the H100 offload sweep."
     - "Preserve the existing DEP4/DEP16 topology, single concurrency-256 point, real model eval, canonical warmup, and one-hour profiling."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Enable recompute for the prefill MultiConnector's optional Mooncake cache loads, using the validation image's tested hybrid recovery when a key disappears after lookup."
+    - "Keep decode-side NIXL transfer failures fail-closed, with unchanged topology, concurrency, eval, warmup, and profiling duration."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From 7187c077abaee172e938aa5061fb64f504e55f03 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 2 Sep 2026 23:53:39 -0500
Subject: [PATCH 16/17] fix(config): route GB300 partition overrides to batch_1

---
 configs/nvidia-master.yaml | 26 +++++++++++++-------------
 perf-changelog.yaml        |  9 +++++++++
 2 files changed, 22 insertions(+), 13 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 4f0a817160..f9296a585e 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8380,7 +8380,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-1p4d-dep4-tep8-c4-b1-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 4
           tp: 8
@@ -8397,7 +8397,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-1p6d-dep4-tep4-c24-b4-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 6
           tp: 4
@@ -8414,7 +8414,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-1p1d-dep8-dep32-c388-b4-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 32
@@ -8431,7 +8431,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-2p1d-dep8-dep32-c736-b8-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 32
@@ -8448,7 +8448,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-3p1d-dep8-dep16-c1152-b32-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 16
@@ -8465,7 +8465,7 @@ dsv4-fp4-gb300-dynamo-trt-agentx:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/deepseek-v4/agentx-gb300-20260811/dynamo-disagg-gb300-5p1d-dep8-dep16-c2626-b96-mtp.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 16
@@ -9235,7 +9235,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-agg-mtp:
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-agg-gb300-tp8-c1-b2-mtp8.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
 # GLM-5.2 NVFP4 TensorRT-LLM AgentX on GB300 with Dynamo disaggregated serving.
 # Six topology variants use NIXL KV transfer and MTP3/MTP5 decoding.
 glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
@@ -9264,7 +9264,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-1p4d-tep4-c30-b2-mtp5.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 4
           tp: 4
@@ -9281,7 +9281,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-1p1d-tep8-c20-b5-mtp5.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 8
@@ -9298,7 +9298,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-3p4d-tep4-c60-b5-mtp5.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 4
           tp: 4
@@ -9315,7 +9315,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-5p1d-dep16-c152-b16-mtp3.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 16
@@ -9332,7 +9332,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-8p1d-dep16-c259-b16-mtp3.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 16
@@ -9349,7 +9349,7 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5.2/gb300-fp4/agentic/dynamo-disagg-gb300-6p1d-dep8-c227-b16-mtp3.yaml"
-          - "SLURM_PARTITION=batch_3"
+          - "SLURM_PARTITION=batch_1"
         decode:
           num-worker: 1
           tp: 8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 97f7baf0df..33982d4e16 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6896,3 +6896,12 @@
     - "Enable recompute for the prefill MultiConnector's optional Mooncake cache loads, using the validation image's tested hybrid recovery when a key disappears after lookup."
     - "Keep decode-side NIXL transfer failures fail-closed, with unchanged topology, concurrency, eval, warmup, and profiling duration."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Standardize all remaining GB300 partition overrides in this validation branch on batch_1, including the DeepSeek-V4 and GLM-5.2 TensorRT-LLM configs. Preserve historical changelog entries."
+    - "Keep the official validation sweep limited to the existing vLLM P/D concurrency-256 point and its full evaluation; its launcher already defaults to batch_1. No image, topology, warmup, or profiling changes."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797

From b41d05864fa0f17674b36b8eb89cfaa21c58468b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 3 Sep 2026 09:47:43 -0500
Subject: [PATCH 17/17] fix(benchmark): use P/D image with complete connector
 block tables

---
 .../disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml   | 4 ++--
 configs/nvidia-master.yaml                               | 2 +-
 perf-changelog.yaml                                      | 9 +++++++++
 3 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
index a03716c7a3..e5d56b3214 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic.yaml
@@ -9,14 +9,14 @@ name: "svf-vllm-disagg-gb300-1p1d-dep4-dep16-c256-mtp-agentic"
 
 model:
   path: "deepseek-v4-pro"
-  container: "ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h"
+  container: "ttl.sh/cquil11-vllm-tier-pd-block-state-0d8739d0-arm64-20260903:24h"
   precision: "fp4"
 
 identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
-    image: "ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h"
+    image: "ttl.sh/cquil11-vllm-tier-pd-block-state-0d8739d0-arm64-20260903:24h"
   frameworks:
     dynamo: "1.4.0"
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f9296a585e..0938a6dcd9 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8049,7 +8049,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg:
           - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml"
 dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
-  image: ttl.sh/cquil11-vllm-tier-pd-recovery-4d3c4a23-arm64-20260903:24h
+  image: ttl.sh/cquil11-vllm-tier-pd-block-state-0d8739d0-arm64-20260903:24h
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 33982d4e16..0fa0f4c34f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6905,3 +6905,12 @@
     - "Standardize all remaining GB300 partition overrides in this validation branch on batch_1, including the DeepSeek-V4 and GLM-5.2 TensorRT-LLM configs. Preserve historical changelog entries."
     - "Keep the official validation sweep limited to the existing vLLM P/D concurrency-256 point and its full evaluation; its launcher already defaults to batch_1. No image, topology, warmup, or profiling changes."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use the 0d8739d0 Python-only validation image with upstream PR #54853 complete connector block tables and PR #54643 rejected Mooncake-load handling, fixing the missing-current-block-table assertion observed in the P/D warmup."
+    - "Preserve the metric implementation, DEP4/DEP16 topology, single concurrency-256 point, full evaluation, canonical warmup, and one-hour profile."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2797