diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-container-deps.sh b/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-container-deps.sh
new file mode 100644
index 0000000000..d11740d596
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-container-deps.sh
@@ -0,0 +1,13 @@
+#!/bin/bash
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+set -euo pipefail
+
+apt-get -y update && apt-get install -y --no-install-recommends --allow-change-held-packages numactl
+
+pip install msgpack
+
+if [ -f /configs/patches/vllm_numa_bind_hash_fix.py ]; then
+    python3 /configs/patches/vllm_numa_bind_hash_fix.py
+fi
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/1k1k/agg-h100-tep8pp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/1k1k/agg-h100-tep8pp2.yaml
new file mode 100644
index 0000000000..c548caf6cb
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/1k1k/agg-h100-tep8pp2.yaml
@@ -0,0 +1,101 @@
+name: "dsv4-vllm-agg-h100-tep8pp2-1k1k"
+
+# DeepSeek-V4-Pro FP8 H100 AGGREGATED TP8 x PP2 recipe (2 nodes / 16 GPUs).
+# The ~1TB FP8 checkpoint does not fit one 8xH100 node (640GB). TP8 x PP2
+# shards attention/dense (/8) and layers (/2) across 16 GPUs -- the original
+# memory-correct sizing that loaded the model and reached prefill. That
+# topology's ONLY failure was the disaggregated NIXL KV-transfer stall between
+# the prefill and decode workers; running AGGREGATED (single worker, decode
+# num-worker 0) removes the P/D split entirely, so there is no NIXL transfer to
+# stall on. TP8 (not TP16) preserves the FP8 128-column block divisibility
+# (moe_intermediate 3072 / 8 = 384, divisible by 128; / 16 = 192 is not).
+# Node allocation for this framework is tp*pp/gpus_per_node = 8*2/8 = 2 nodes;
+# EP and DP-attention do NOT contribute GPUs, which is why the earlier DP32
+# aggregated attempt collapsed onto a single node.
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:v0.21.0"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 1440
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+  exclude: "hpc-gpu-1-0,hpc-gpu-1-1,hpc-gpu-1-4,hpc-gpu-1-5,hpc-gpu-1-7,hpc-gpu-1-8,hpc-gpu-1-13,hpc-gpu-1-16,hpc-gpu-1-19"
+
+resources:
+  gpu_type: "h100"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    NCCL_CUMEM_ENABLE: "1"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "all"
+    VLLM_SERVER_DEV_MODE: "1"
+  vllm_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      block-size: 256
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 2
+      enable-expert-parallel: true
+      kv-cache-dtype: "fp8"
+      gpu-memory-utilization: 0.92
+      max-model-len: 2304
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      # CUDA graphs (not eager): the eager run served correctly but decode ran
+      # at ~144ms/token (Python dispatch-bound, flat vs concurrency), so the
+      # conc sweep could not finish in 8h. On-cluster prototype (job 11256)
+      # confirmed FULL_AND_PIECEWISE graphs FIT this 2-node TP8xPP2 layout at
+      # gpu-mem-util 0.92 and drop TPOT to ~15ms (~9.4x). No extra nodes needed.
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      reasoning-parser: deepseek_v4
+      tokenizer-mode: deepseek_v4
+      trust-remote-code: true
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "1x2x4x8x16x32x64x128x256x512"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+    revision: "0366e4e064385807ea86b088a5c6c878ff23343b"
+  container:
+    image: "vllm/vllm-openai:v0.21.0"
+  frameworks:
+    dynamo: "1.2.0.dev20260426"
+    vllm: "0.21.0"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-h100-tep8pp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-h100-tep8pp2.yaml
new file mode 100644
index 0000000000..b55cdc4b13
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-h100-tep8pp2.yaml
@@ -0,0 +1,103 @@
+name: "dsv4-vllm-agg-h100-tep8pp2-8k1k"
+
+# DeepSeek-V4-Pro FP8 H100 AGGREGATED TP8 x PP2 recipe (2 nodes / 16 GPUs).
+# The ~1TB FP8 checkpoint does not fit one 8xH100 node (640GB). TP8 x PP2
+# shards attention/dense (/8) and layers (/2) across 16 GPUs -- the original
+# memory-correct sizing that loaded the model and reached prefill. That
+# topology's ONLY failure was the disaggregated NIXL KV-transfer stall between
+# the prefill and decode workers; running AGGREGATED (single worker, decode
+# num-worker 0) removes the P/D split entirely, so there is no NIXL transfer to
+# stall on. TP8 (not TP16) preserves the FP8 128-column block divisibility
+# (moe_intermediate 3072 / 8 = 384, divisible by 128; / 16 = 192 is not).
+# Node allocation for this framework is tp*pp/gpus_per_node = 8*2/8 = 2 nodes;
+# EP and DP-attention do NOT contribute GPUs, which is why the earlier DP32
+# aggregated attempt collapsed onto a single node.
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:v0.21.0"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  interval_seconds: 10
+  max_attempts: 1440
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+  exclude: "hpc-gpu-1-0,hpc-gpu-1-1,hpc-gpu-1-4,hpc-gpu-1-5,hpc-gpu-1-7,hpc-gpu-1-8,hpc-gpu-1-13,hpc-gpu-1-16,hpc-gpu-1-19"
+
+resources:
+  gpu_type: "h100"
+  gpus_per_node: 8
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 16
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    NCCL_CUMEM_ENABLE: "1"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: "all"
+    VLLM_SERVER_DEV_MODE: "1"
+  vllm_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      block-size: 256
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 2
+      enable-expert-parallel: true
+      kv-cache-dtype: "fp8"
+      gpu-memory-utilization: 0.92
+      max-model-len: 9472
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      # CUDA graphs (not eager): the eager run served correctly but decode ran
+      # at ~144ms/token (Python dispatch-bound, flat vs concurrency), so the
+      # conc sweep could not finish in 8h. On-cluster prototype (job 11256)
+      # confirmed FULL_AND_PIECEWISE graphs FIT this 2-node TP8xPP2 layout at
+      # gpu-mem-util 0.92 and drop TPOT to ~15ms (~9.4x). No extra nodes needed.
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      reasoning-parser: deepseek_v4
+      tokenizer-mode: deepseek_v4
+      trust-remote-code: true
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  # Capped at 256: conc512 (5120 x 8k-token requests) needs ~3.6h for that one
+  # level and exceeds the 8h SLURM limit; conc256 completes (~71min) and fits.
+  concurrencies: "1x2x4x8x16x32x64x128x256"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+    revision: "0366e4e064385807ea86b088a5c6c878ff23343b"
+  container:
+    image: "vllm/vllm-openai:v0.21.0"
+  frameworks:
+    dynamo: "1.2.0.dev20260426"
+    vllm: "0.21.0"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f90c510186..3d99512a8a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -4588,6 +4588,80 @@ dsr1-fp8-h100-dynamo-sglang:
           ep: 16
           dp-attn: true
 
+# DeepSeek-V4-Pro FP8 H100 aggregated vLLM via Dynamo (TP8 x PP2, 2 nodes).
+# The native checkpoint is mixed FP4+FP8 (~1TB), too large for one 8xH100 node
+# (640GB). Prior attempts: TP8xPP2 *disagg* loaded the model and reached prefill
+# but hung on the NIXL P/D KV-transfer; DP16/2-node OOM'd; DP32 "aggregated"
+# collapsed onto ONE node because this framework sizes nodes as tp*pp/gpus_per_node
+# (EP and DP-attention add no GPUs), so tp1 => 1 node and vLLM could not place 32
+# DP ranks on 8 GPUs. The working topology is the memory-proven TP8xPP2 run made
+# AGGREGATED (single worker, no P/D split => no NIXL transfer to stall on): tp8*pp2
+# = 16 GPUs = 2 nodes. TP8 (not TP16) keeps FP8 128-column block divisibility.
+# Recipes: benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/{1k1k,8k1k}/agg-h100-tep8pp2.yaml.
+dsv4-fp8-h100-dynamo-vllm:
+  image: vllm/vllm-openai:v0.21.0
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  # Runs-on label = this string verbatim. Must be a label a live runner
+  # carries: h100-disagg has 3 online runners (h100-dgxc-slurm_17-19) and is
+  # the disaggregated-H100 pool. NOT h100-multinode — no runner carries that
+  # label, so jobs requesting it queue forever (the dsr1-fp8-h100-dynamo-*
+  # reference configs above still use it and are non-functional for this).
+  runner: h100-disagg
+  precision: fp8
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.2.0.dev20260426" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  # Aggregated (not P/D-disaggregated): prefill num-worker 1 + decode
+  # num-worker 0 is how srtslurm signals AGG mode (see RECIPES.md §5). Node
+  # count = tp*pp/gpus_per_node = 8*2/8 = 2. tp:8 pp:2 with expert-parallel;
+  # dp-attn false (TP8 attention). Aggregated => no NIXL transfer to stall on.
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # Aggregated TP8 x PP2 over 2 H100 nodes (16 GPUs), EP8.
+      - conc-list: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512]
+        prefill:
+          num-worker: 1
+          tp: 8
+          pp: 2
+          ep: 8
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/1k1k/agg-h100-tep8pp2.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          pp: 2
+          ep: 8
+          dp-attn: false
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # Aggregated TP8 x PP2 over 2 H100 nodes (16 GPUs), EP8.
+      # 8k1k capped at conc256: conc512 (5120 x 8k-token requests) needs ~3.6h
+      # for that single level and blew the 8h SLURM limit; conc256 completes in
+      # ~71min and fits with margin. 1k1k keeps 512 (short requests finish fast).
+      - conc-list: [1, 2, 4, 8, 16, 32, 64, 128, 256]
+        prefill:
+          num-worker: 1
+          tp: 8
+          pp: 2
+          ep: 8
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/agg-h100-tep8pp2.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          pp: 2
+          ep: 8
+          dp-attn: false
+
 minimaxm3-fp8-h200-vllm:
   image: vllm/vllm-openai:minimax-m3
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/configs/runners.yaml b/configs/runners.yaml
index 93da791b85..135c8fb9e8 100644
--- a/configs/runners.yaml
+++ b/configs/runners.yaml
@@ -27,6 +27,13 @@ labels:
     - h100-dgxc-slurm_17
     - h100-dgxc-slurm_18
     - h100-dgxc-slurm_19
+  # Disaggregated-H100 pool. Unlike h100-multinode (whose key is not a label
+  # any runner carries, so jobs requesting it queue forever), these three
+  # runners carry the h100-disagg label, so runs-on: h100-disagg matches.
+  h100-disagg:
+    - h100-dgxc-slurm_17
+    - h100-dgxc-slurm_18
+    - h100-dgxc-slurm_19
   h200:
     - h200-cw_00
     - h200-cw_01
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e05c2ca1c4..b92f24cd37 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4751,6 +4751,22 @@
     - "6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2137
 
+- config-keys:
+    - dsv4-fp8-h200-vllm
+    - dsv4-fp8-h200-vllm-mtp
+  description:
+    - "Bump vLLM image from v0.21.0 to v0.25.0 for DeepSeek-V4-Pro FP8 on H200, matching the B200/B300 dsv4 vLLM bump (#2169)"
+    - "Refresh stale H200 dsv4 submissions (last run 2026-05-21)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2191
+
+- config-keys:
+    - dsv4-fp8-h200-sglang
+    - dsv4-fp8-h200-sglang-mtp
+  description:
+    - "Bump the pinned lmsysorg/sglang:deepseek-v4-hopper digest from the 2026-05-02 push (7f19c6dc) to the current 2026-05-13 push (1bf5d508)"
+    - "Refresh stale H200 dsv4 SGLang submissions (last run 2026-05-04)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2191
+
 - config-keys:
     - dsv4-fp4-mi355x-sglang
   description:
@@ -4765,3 +4781,18 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708"
     - "Clean the export envs"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2198
+
+- config-keys:
+    - dsv4-fp8-h100-dynamo-vllm
+  description:
+    - "Add DeepSeek-V4-Pro FP8 H100 aggregated multinode vLLM benchmark via Dynamo (new SKU, previously no dsv4 data on H100)"
+    - "Native checkpoint is mixed FP4+FP8 (~1TB) and does not fit one 8xH100 node (640GB); vLLM auto-resolves the DSv4 quant config to a Hopper-compatible path (no explicit --quantization), so the only real blocker on H100 is memory, not FP4 support"
+    - "Aggregated TP8 x PP2 across 2 H100 nodes (16 GPUs) with expert-parallel: tensor-parallel-size 8, pipeline-parallel-size 2, enable-expert-parallel. TP8 shards attention/dense (/8) and PP2 splits layers (/2) -- the original memory-correct sizing that loaded the model and reached prefill. TP8 (not TP16) preserves the FP8 128-column quantization-block divisibility (moe_intermediate 3072/8=384 divisible by 128; /16=192 not)"
+    - "Native checkpoint quant note: vLLM auto-resolves the DSv4 config to a Hopper-compatible path (no explicit --quantization); the only blocker on H100 is memory, not FP4 support (all failed attempts loaded far enough to fail on placement/transfer, not on FP4)"
+    - "Supersedes three failed bring-up topologies: (1) TP8xPP2 *disaggregated* loaded the model and reached prefill but hung ~300s on the NIXL P/D KV-transfer then EngineCore died; (2) pure DP16 on 2 nodes OOM'd at model load; (3) DP32 'aggregated' collapsed onto ONE node -- this framework sizes nodes as tp*pp/gpus_per_node and EP/DP-attention add no GPUs, so tp1 => 1 node and vLLM could not place 32 DP ranks on 8 GPUs. Running the memory-proven TP8xPP2 AGGREGATED (single worker, no P/D split) removes the NIXL transfer entirely and allocates 2 nodes deterministically (8*2/8)"
+    - "Aggregated mode is signalled by prefill num-worker 1 + decode num-worker 0 (RECIPES.md section 5); the full vLLM serve config lives under backend.vllm_config.aggregated in the recipe"
+    - "CUDA graphs (compilation-config cudagraph_mode FULL_AND_PIECEWISE, custom_ops all) at gpu-memory-utilization 0.92 -- NOT enforce-eager. An eager run served correctly (eval passed) but decode ran at ~144ms/token (Python-dispatch-bound, flat vs concurrency), so the conc sweep could not finish in the 8h SLURM limit (killed mid-conc-128 after ~7h). An on-cluster srtctl prototype (2-node TP8xPP2, 1x8 sweep) confirmed FULL_AND_PIECEWISE graphs FIT this layout (no OOM at 0.92) and cut TPOT to ~15ms (~9.4x); no extra nodes required"
+    - "One TP8xPP2 topology per seq-len; 1k1k conc 1-512, 8k1k conc 1-256 (8k1k conc512 = 5120 x 8k-token requests needs ~3.6h for that single level and exceeds the 8h SLURM limit; conc256 completes in ~71min). FP8 KV cache, block size 256, max-model-len 2304/9472, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True"
+    - "Recipes under benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/{1k1k,8k1k}/agg-h100-tep8pp2.yaml; setup via NVIDIA srt-slurm's standard vllm-container-deps.sh"
+    - "Runner pool: h100-disagg (3 online runners), not the h100-multinode label used by the dsr1-fp8-h100-dynamo-* references — no runner carries that label, so jobs requesting it queue indefinitely"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2196
diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index 1334c95542..29aeab9330 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -16,7 +16,30 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     # MODEL_PATH: Override with pre-downloaded paths on H100 runner
     # The yaml files specify HuggingFace model IDs for portability, but we use
     # local paths to avoid repeated downloading on the shared H100 cluster.
-    if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
+    if [[ $FRAMEWORK == "dynamo-vllm" ]]; then
+        if [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp8" ]]; then
+            SELECTED_MODEL_PATH=""
+            if [[ -n "${MODEL_PATH:-}" && -d "${MODEL_PATH}" ]]; then
+                SELECTED_MODEL_PATH="$MODEL_PATH"
+            else
+                for candidate in \
+                    /mnt/nfs/lustre/models/dsv4-fp8 \
+                    /mnt/nfs/lustre/models/deepseek-v4-pro \
+                    /mnt/nfs/lustre/models/dsv4-pro \
+                    /mnt/nfs/lustre/models/DeepSeek-V4-Pro; do
+                    if [[ -d "$candidate" ]]; then
+                        SELECTED_MODEL_PATH="$candidate"
+                        break
+                    fi
+                done
+            fi
+            export MODEL_PATH="${SELECTED_MODEL_PATH:-/mnt/nfs/lustre/models/dsv4-fp8}"
+            export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
+        else
+            echo "Unsupported model prefix/precision for dynamo-vllm: $MODEL_PREFIX/$PRECISION"
+            exit 1
+        fi
+    elif [[ $FRAMEWORK == "dynamo-sglang" ]]; then
         if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
             export MODEL_PATH="/mnt/nfs/lustre/models/dsr1-fp8"
             export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
@@ -34,7 +57,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
             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
 
@@ -49,6 +72,15 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     if [[ "$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"
+    elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
+        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+        git checkout aflowers/vllm-gb200-v0.20.0
+        mkdir -p recipes/vllm/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+        cp \
+            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-container-deps.sh" \
+            configs/patches/vllm-container-deps.sh
     else
         git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR"
@@ -84,6 +116,23 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         # SGLang container mapping
         SQUASH_FILE="/mnt/nfs/lustre/containers/lmsysorg_sglang_v0.5.8.post1-cu130.sqsh"
         CONTAINER_KEY="lmsysorg/sglang:v0.5.8-cu130"
+    elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
+        # vLLM container mapping. Import the recipe image once into the shared
+        # squash directory so every Slurm node uses the exact configured tag.
+        CONTAINER_KEY="$IMAGE"
+        SQUASH_FILE="/mnt/nfs/sa-shared/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/+/g').sqsh"
+        mkdir -p "$(dirname "$SQUASH_FILE")"
+        (
+            exec 9>"${SQUASH_FILE}.lock"
+            flock -w 1800 9 || { echo "Failed to acquire lock for $SQUASH_FILE" >&2; exit 1; }
+            if unsquashfs -l "$SQUASH_FILE" >/dev/null 2>&1; then
+                echo "Squash file already exists and is valid: $SQUASH_FILE"
+            else
+                rm -f "$SQUASH_FILE" "${SQUASH_FILE}.tmp."*
+                enroot import -o "${SQUASH_FILE}.tmp.$$" "docker://$IMAGE"
+                mv -f "${SQUASH_FILE}.tmp.$$" "$SQUASH_FILE"
+            fi
+        )
     elif [[ $FRAMEWORK == "dynamo-trt" ]]; then
         # TRT-LLM container mapping - convert IMAGE to srt-slurm format (nvcr.io/ -> nvcr.io#)
         CONTAINER_KEY=$(echo "$IMAGE" | sed 's|nvcr.io/|nvcr.io#|')
@@ -115,6 +164,7 @@ model_paths:
 containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
+  dynamo-vllm: "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
   latest: "${SQUASH_FILE}"
   "${CONTAINER_KEY}": "${SQUASH_FILE}"
@@ -143,8 +193,10 @@ EOF
 
     # Override the job name in the config file with the runner name
     sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
-    # Raise sglang's torch-distributed TCPStore timeout from the 600s gloo default
-    sed -i '/^      watchdog-timeout:/a\      dist-timeout: 1800' "${CONFIG_FILE%%:*}"
+    # Raise sglang's torch-distributed TCPStore timeout from the 600s gloo default.
+    if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
+        sed -i '/^      watchdog-timeout:/a\      dist-timeout: 1800' "${CONFIG_FILE%%:*}"
+    fi
     SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "h100,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
     echo "$SRTCTL_OUTPUT"