diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep16-conc8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep16-conc8192.yaml
new file mode 100644
index 0000000000..021e430ebe
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep16-conc8192.yaml
@@ -0,0 +1,200 @@
+# 1k1k 1P1D / DEP16 / total 5 nodes
+# conc sweep: 2048, 4096, 8192 (c/4, c/2, c where c = X * 8192)
+# stack: lmsysorg/sglang nightly-dev-cu13-20260707-b4155233 (megamoe) + dynamo 81d0555e
+
+name: "disagg-1p1d-dep16-conc8192"
+
+slurm:
+  time_limit: "03:00:00"
+
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+
+# Backend workers build dynamo from source; that install runs
+# `apt-get install libclang-dev`, which needs root. Run the worker container
+# as remapped root so dpkg works (pyxis maps container-root back to the
+# invoking user on the host, so files on mounted dirs stay caller-owned).
+# Without this, workers run non-root and the install dies with exit 100
+# (`dpkg: requested operation requires superuser privilege`), hidden by the
+# install's `> /dev/null 2>&1`.
+srun_options:
+  container-remap-root: ""
+  # Step-level --mem=0: give each worker srun step all node RAM so the
+  # DSV4-Pro weight load does not hit the step cgroup limit and stall/OOM
+  # (job-level `sbatch --mem=0` does not lift the per-step memory cgroup).
+  mem: "0"
+
+dynamo:
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
+  install: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+  env:
+    DYN_ROUTER_LOAD_BLOCK_SIZE: "1"
+  args:
+    router-mode: "kv"
+    router-kv-overlap-score-weight: 0
+    router-queue-threshold: 64
+    router-temperature: 0.5
+    no-kv-events: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: fp4
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  decode_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '1152'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: "8"
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  sglang_config:
+    prefill:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      moe-a2a-backend: megamoe
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 16
+      data-parallel-size: 16
+      expert-parallel-size: 16
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.94
+      swa-full-tokens-ratio: 0.12
+      context-length: 16384
+      max-running-requests: 18432
+      cuda-graph-max-bs: 1152
+
+benchmark:
+  type: custom
+  command: |
+      set -e
+      ulimit -n 1048576
+      REPO=/configs/upstream-sa-bench/InferenceX
+      [ -d "$REPO" ] || git clone https://github.com/SemiAnalysisAI/InferenceX.git "$REPO"
+      cd "$REPO/utils/bench_serving"
+
+      echo "=== block conc=2048 prompts=6144 warmups=2048 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 6144 --max-concurrency 2048 --request-rate inf \
+        --num-warmups 2048 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc2048.json
+
+      echo "=== block conc=4096 prompts=12288 warmups=4096 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 12288 --max-concurrency 4096 --request-rate inf \
+        --num-warmups 4096 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc4096.json
+
+      echo "=== block conc=8192 prompts=24576 warmups=8192 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 24576 --max-concurrency 8192 --request-rate inf \
+        --num-warmups 8192 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc8192.json
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep8-conc8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep8-conc8192.yaml
new file mode 100644
index 0000000000..2232e6f817
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep8-conc8192.yaml
@@ -0,0 +1,201 @@
+# 1k1k 1P1D / DEP8 / total 3 nodes
+# conc sweep: 2048, 4096, 8192 (c/4, c/2, c where c = X * 8192)
+# stack: lmsysorg/sglang nightly-dev-cu13-20260707-b4155233 (megamoe) + dynamo 81d0555e
+
+name: "disagg-1p1d-dep8-conc8192"
+
+slurm:
+  time_limit: "03:00:00"
+
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+
+# Backend workers build dynamo from source; that install runs
+# `apt-get install libclang-dev`, which needs root. Run the worker container
+# as remapped root so dpkg works (pyxis maps container-root back to the
+# invoking user on the host, so files on mounted dirs stay caller-owned).
+# Without this, workers run non-root and the install dies with exit 100
+# (`dpkg: requested operation requires superuser privilege`), hidden by the
+# install's `> /dev/null 2>&1`.
+srun_options:
+  container-remap-root: ""
+  # Step-level --mem=0: give each worker srun step all node RAM so the
+  # DSV4-Pro weight load does not hit the step cgroup limit and stall/OOM
+  # (job-level `sbatch --mem=0` does not lift the per-step memory cgroup).
+  mem: "0"
+
+dynamo:
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
+  install: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+  env:
+    DYN_ROUTER_LOAD_BLOCK_SIZE: "1"
+  args:
+    router-mode: "kv"
+    router-kv-overlap-score-weight: 0
+    router-queue-threshold: 64
+    router-temperature: 0.5
+    no-kv-events: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: fp4
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 2
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  decode_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '1152'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: "8"
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  sglang_config:
+    prefill:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      moe-a2a-backend: megamoe
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.94
+      swa-full-tokens-ratio: 0.12
+      context-length: 16384
+      max-running-requests: 9216
+      cuda-graph-max-bs: 1152
+
+benchmark:
+  type: custom
+  command: |
+      set -e
+      ulimit -n 1048576
+      REPO=/configs/upstream-sa-bench/InferenceX
+      [ -d "$REPO" ] || git clone https://github.com/SemiAnalysisAI/InferenceX.git "$REPO"
+      cd "$REPO/utils/bench_serving"
+
+      echo "=== block conc=2048 prompts=6144 warmups=2048 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 6144 --max-concurrency 2048 --request-rate inf \
+        --num-warmups 2048 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc2048.json
+
+      echo "=== block conc=4096 prompts=12288 warmups=4096 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 12288 --max-concurrency 4096 --request-rate inf \
+        --num-warmups 4096 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc4096.json
+
+      echo "=== block conc=8192 prompts=24576 warmups=8192 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 24576 --max-concurrency 8192 --request-rate inf \
+        --num-warmups 8192 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc8192.json
+
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p4d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p4d-dep4-tp4.yaml
new file mode 100644
index 0000000000..3ec6102c25
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p4d-dep4-tp4.yaml
@@ -0,0 +1,127 @@
+name: "dsv4-pro-gb300-disagg-1k1k-1p4d-dep4-tp4"
+
+# DeepSeek-V4-Pro 1k/1k disagg on GB300 — 1P + 4D, DEP4 prefill / TP4 decode.
+#
+# Single-config recipe derived from disagg-low-latency-dep4.yaml zip override
+# variant [2] (decode_nodes=4). Concurrencies pruned to the Pareto-frontier
+# points from the original sweep.
+
+# GB300 host-RAM: give the job all node memory and each srun step --mem=0
+# so the DSV4-Pro weight load is not constrained by the per-step memory
+# cgroup (job-level mem=0 alone does not lift it). Matches minimax GB300.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+srun_options:
+  mem: "0"
+
+frontend:
+  type: sglang
+  enable_multiple_frontends: false
+  args:
+    policy: "cache_aware"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: "mxfp4"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 4
+  decode_workers: 4
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "prefill"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+
+      moe-a2a-backend: "megamoe"
+
+      mem-fraction-static: 0.90
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 4096
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "decode"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size:   1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      mem-fraction-static: 0.9
+      max-running-requests: 128
+      cuda-graph-max-bs:    128
+      swa-full-tokens-ratio: 0.1
+      context-length: 4096
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "8x64"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p6d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p6d-dep4-tp4.yaml
new file mode 100644
index 0000000000..6a8767baae
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-1p6d-dep4-tp4.yaml
@@ -0,0 +1,126 @@
+name: "dsv4-pro-gb300-disagg-1k1k-1p6d-dep4-tp4"
+# DeepSeek-V4-Pro 1k/1k disagg on GB300 — 1P + 6D, DEP4 prefill / TP4 decode.
+#
+# Single-config recipe derived from disagg-low-latency-dep4.yaml zip override
+# variant [3] (decode_nodes=6). Concurrencies pruned to the Pareto-frontier
+# points from the original sweep
+
+# GB300 host-RAM: give the job all node memory and each srun step --mem=0
+# so the DSV4-Pro weight load is not constrained by the per-step memory
+# cgroup (job-level mem=0 alone does not lift it). Matches minimax GB300.
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+srun_options:
+  mem: "0"
+
+frontend:
+  type: sglang
+  enable_multiple_frontends: false
+  args:
+    policy: "cache_aware"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: "mxfp4"
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  decode_nodes: 6
+  decode_workers: 6
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "0"
+    SGLANG_ENABLE_THINKING: "1"
+    SGLANG_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "prefill"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+
+      moe-a2a-backend: "megamoe"
+
+      mem-fraction-static: 0.90
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 4096
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      disable-radix-cache: true
+
+      disaggregation-mode: "decode"
+      disaggregation-transfer-backend: mooncake
+
+      tensor-parallel-size: 4
+      data-parallel-size:   1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      mem-fraction-static: 0.9
+      max-running-requests: 128
+      cuda-graph-max-bs:    128
+      swa-full-tokens-ratio: 0.1
+      context-length: 4096
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "32x64"
+  req_rate: "inf"
+  use_chat_template: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep12-conc18432.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep12-conc18432.yaml
new file mode 100644
index 0000000000..ef78f369b3
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep12-conc18432.yaml
@@ -0,0 +1,144 @@
+# 1k1k 2P1D / DEP12 / total 5 nodes
+# rate-match point between 2:4 and 3:4; c8bd stack, single conc=18432, mrr=13824
+# stack: lmsysorg/sglang nightly-dev-cu13-20260707-b4155233 (megamoe) + dynamo 81d0555e
+
+name: "disagg-2p1d-dep12-conc18432"
+slurm:
+  time_limit: 03:00:00
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+# Backend workers build dynamo from source; that install runs
+# `apt-get install libclang-dev`, which needs root. Run the worker container
+# as remapped root so dpkg works (pyxis maps container-root back to the
+# invoking user on the host, so files on mounted dirs stay caller-owned).
+# Without this, workers run non-root and the install dies with exit 100
+# (`dpkg: requested operation requires superuser privilege`), hidden by the
+# install's `> /dev/null 2>&1`.
+srun_options:
+  container-remap-root: ""
+  # Step-level --mem=0: give each worker srun step all node RAM so the
+  # DSV4-Pro weight load does not hit the step cgroup limit and stall/OOM
+  # (job-level `sbatch --mem=0` does not lift the per-step memory cgroup).
+  mem: "0"
+
+dynamo:
+  hash: 81d0555ee23519cea80a42b4fe824e30368b7300
+  install: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+  env:
+    DYN_ROUTER_LOAD_BLOCK_SIZE: '1'
+  args:
+    router-mode: kv
+    router-kv-overlap-score-weight: 0
+    router-queue-threshold: 64
+    router-temperature: 0.5
+    no-kv-events: true
+model:
+  path: "deepseek-v4-pro"
+  container: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+  decode_nodes: 3
+  decode_workers: 1
+  gpus_per_decode: 12
+backend:
+  type: sglang
+  prefill_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '8192'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+  decode_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '1152'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: '1'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+  sglang_config:
+    prefill:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      moe-a2a-backend: megamoe
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+    decode:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 12
+      data-parallel-size: 12
+      expert-parallel-size: 12
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.94
+      swa-full-tokens-ratio: 0.12
+      context-length: 16384
+      max-running-requests: 13824
+      cuda-graph-max-bs: 1152
+benchmark:
+  type: custom
+  command: "set -e\nulimit -n 1048576\nREPO=/configs/upstream-sa-bench/InferenceX\n[ -d \"$REPO\" ] || git clone https://github.com/SemiAnalysisAI/InferenceX.git \"$REPO\"\ncd \"$REPO/utils/bench_serving\"\n\necho \"=== 2P1D DEP12 rate-match point: conc=18432 prompts=55296 warmups=18432 mrr=13824 ===\"\npython3 benchmark_serving.py \\\n  --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \\\n  --host 127.0.0.1 --port 8000 --endpoint /v1/completions \\\n  --dataset-name random \\\n  --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \\\n  --random-num-workers 96 \\\n  --num-prompts 55296 --max-concurrency 18432 --request-rate inf \\\n  --num-warmups 18432 \\\n  --ignore-eos --trust-remote-code \\\n  --percentile-metrics ttft,tpot,itl,e2el \\\n  --save-result --result-dir /logs --result-filename results_conc18432_mrr13824.json\n"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep16-conc16384.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep16-conc16384.yaml
new file mode 100644
index 0000000000..2826b7d9a7
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep16-conc16384.yaml
@@ -0,0 +1,200 @@
+# 1k1k 2P1D / DEP16 / total 6 nodes
+# conc sweep: 4096, 8192, 16384 (c/4, c/2, c where c = X * 8192)
+# stack: lmsysorg/sglang nightly-dev-cu13-20260707-b4155233 (megamoe) + dynamo 81d0555e
+
+name: "disagg-2p1d-dep16-conc16384"
+
+slurm:
+  time_limit: "03:00:00"
+
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+
+# Backend workers build dynamo from source; that install runs
+# `apt-get install libclang-dev`, which needs root. Run the worker container
+# as remapped root so dpkg works (pyxis maps container-root back to the
+# invoking user on the host, so files on mounted dirs stay caller-owned).
+# Without this, workers run non-root and the install dies with exit 100
+# (`dpkg: requested operation requires superuser privilege`), hidden by the
+# install's `> /dev/null 2>&1`.
+srun_options:
+  container-remap-root: ""
+  # Step-level --mem=0: give each worker srun step all node RAM so the
+  # DSV4-Pro weight load does not hit the step cgroup limit and stall/OOM
+  # (job-level `sbatch --mem=0` does not lift the per-step memory cgroup).
+  mem: "0"
+
+dynamo:
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
+  install: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+  env:
+    DYN_ROUTER_LOAD_BLOCK_SIZE: "1"
+  args:
+    router-mode: "kv"
+    router-kv-overlap-score-weight: 0
+    router-queue-threshold: 64
+    router-temperature: 0.5
+    no-kv-events: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: fp4
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  decode_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '1152'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: "8"
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  sglang_config:
+    prefill:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      moe-a2a-backend: megamoe
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 16
+      data-parallel-size: 16
+      expert-parallel-size: 16
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.94
+      swa-full-tokens-ratio: 0.12
+      context-length: 16384
+      max-running-requests: 18432
+      cuda-graph-max-bs: 1152
+
+benchmark:
+  type: custom
+  command: |
+      set -e
+      ulimit -n 1048576
+      REPO=/configs/upstream-sa-bench/InferenceX
+      [ -d "$REPO" ] || git clone https://github.com/SemiAnalysisAI/InferenceX.git "$REPO"
+      cd "$REPO/utils/bench_serving"
+
+      echo "=== block conc=4096 prompts=12288 warmups=4096 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 12288 --max-concurrency 4096 --request-rate inf \
+        --num-warmups 4096 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc4096.json
+
+      echo "=== block conc=8192 prompts=24576 warmups=8192 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 24576 --max-concurrency 8192 --request-rate inf \
+        --num-warmups 8192 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc8192.json
+
+      echo "=== block conc=16384 prompts=49152 warmups=16384 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 49152 --max-concurrency 16384 --request-rate inf \
+        --num-warmups 16384 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc16384.json
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep48-conc16384.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep48-conc16384.yaml
new file mode 100644
index 0000000000..f5e56890af
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep48-conc16384.yaml
@@ -0,0 +1,202 @@
+# 1k1k 2P1D / DEP48 / total 14 nodes
+# conc sweep: 4096, 8192, 16384 (c/4, c/2, c where c = X * 8192)
+# stack: lmsysorg/sglang nightly-dev-cu13-20260707-b4155233 (megamoe) + dynamo 81d0555e
+
+name: "disagg-2p1d-dep48-conc16384"
+
+slurm:
+  time_limit: "03:00:00"
+
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+
+# Backend workers build dynamo from source; that install runs
+# `apt-get install libclang-dev`, which needs root. Run the worker container
+# as remapped root so dpkg works (pyxis maps container-root back to the
+# invoking user on the host, so files on mounted dirs stay caller-owned).
+# Without this, workers run non-root and the install dies with exit 100
+# (`dpkg: requested operation requires superuser privilege`), hidden by the
+# install's `> /dev/null 2>&1`.
+srun_options:
+  container-remap-root: ""
+  # Step-level --mem=0: give each worker srun step all node RAM so the
+  # DSV4-Pro weight load does not hit the step cgroup limit and stall/OOM
+  # (job-level `sbatch --mem=0` does not lift the per-step memory cgroup).
+  mem: "0"
+
+dynamo:
+  hash: "81d0555ee23519cea80a42b4fe824e30368b7300"
+  install: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+  env:
+    DYN_ROUTER_LOAD_BLOCK_SIZE: "1"
+  args:
+    router-mode: "kv"
+    router-kv-overlap-score-weight: 0
+    router-queue-threshold: 64
+    router-temperature: 0.5
+    no-kv-events: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  precision: fp4
+
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+  decode_nodes: 12
+  decode_workers: 1
+  gpus_per_decode: 48
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  decode_environment:
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+    PYTHONUNBUFFERED: '1'
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '0'
+    SGLANG_ENABLE_THINKING: '1'
+    SGLANG_REASONING_EFFORT: max
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+    SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN: '1'
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '1152'
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "1"
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_CUMEM_ENABLE: '1'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: "8"
+    MC_FORCE_MNNVL: '1'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+    SGLANG_LOG_FORWARD_ITERS: '1'
+    SGLANG_LOG_MS: '1'
+    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+    DYN_LOG: 'info,dynamo_runtime::pipeline::network::ingress::push_handler=warn'
+
+  sglang_config:
+    prefill:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      moe-dense-tp-size: 1
+      enable-dp-lm-head: true
+      moe-a2a-backend: megamoe
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 86400
+      skip-tokenizer-init: true
+      stream-interval: 60
+      tensor-parallel-size: 48
+      data-parallel-size: 48
+      expert-parallel-size: 48
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      mem-fraction-static: 0.94
+      swa-full-tokens-ratio: 0.12
+      context-length: 16384
+      max-running-requests: 55296
+      cuda-graph-max-bs: 1152
+
+benchmark:
+  type: custom
+  command: |
+      set -e
+      ulimit -n 1048576
+      REPO=/configs/upstream-sa-bench/InferenceX
+      [ -d "$REPO" ] || git clone https://github.com/SemiAnalysisAI/InferenceX.git "$REPO"
+      cd "$REPO/utils/bench_serving"
+
+      echo "=== block conc=4096 prompts=12288 warmups=4096 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 12288 --max-concurrency 4096 --request-rate inf \
+        --num-warmups 4096 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc4096.json
+
+      echo "=== block conc=8192 prompts=24576 warmups=8192 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 24576 --max-concurrency 8192 --request-rate inf \
+        --num-warmups 8192 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc8192.json
+
+      echo "=== block conc=16384 prompts=49152 warmups=16384 ==="
+      python3 benchmark_serving.py \
+        --backend vllm --model deepseek-ai/DeepSeek-V4-Pro --tokenizer /model \
+        --host 127.0.0.1 --port 8000 --endpoint /v1/completions \
+        --dataset-name random \
+        --random-input-len 1024 --random-output-len 1024 --random-range-ratio 0.8 \
+        --random-num-workers 96 \
+        --num-prompts 49152 --max-concurrency 16384 --request-rate inf \
+        --num-warmups 16384 \
+        --ignore-eos --trust-remote-code \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --save-result --result-dir /logs --result-filename results_conc16384.json
+
+
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 8d84bd0dd7..fd2feb0ade 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -10173,6 +10173,108 @@ dsv4-fp4-gb300-dynamo-sglang:
           tp: 40
           ep: 40
           dp-attn: true
+    # 1k1k STP disagg configs.
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # 1p1d-dep8. 3 nodes.
+      - conc-list: [8192]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep8-conc8192.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 1p1d-dep16. 5 nodes.
+      - conc-list: [8192]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-1p1d-dep16-conc8192.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      # 1p4d-dep4-tp4. 5 nodes. sa-bench conc 8x64.
+      - conc-list: [64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-1p4d-dep4-tp4.yaml"
+        decode:
+          num-worker: 4
+          tp: 4
+          ep: 1
+          dp-attn: false
+      # 1p6d-dep4-tp4. 7 nodes. sa-bench conc 32x64.
+      - conc-list: [64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-1p6d-dep4-tp4.yaml"
+        decode:
+          num-worker: 6
+          tp: 4
+          ep: 1
+          dp-attn: false
+      # 2p1d-dep12. 5 nodes.
+      - conc-list: [18432]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep12-conc18432.yaml"
+        decode:
+          num-worker: 1
+          tp: 12
+          ep: 12
+          dp-attn: true
+      # 2p1d-dep16. 6 nodes.
+      - conc-list: [16384]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep16-conc16384.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      # 2p1d-dep48. 14 nodes.
+      - conc-list: [16384]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/1k1k/disagg-2p1d-dep48-conc16384.yaml"
+        decode:
+          num-worker: 1
+          tp: 48
+          ep: 48
+          dp-attn: true
 
 glm5-fp8-b200-dynamo-sglang:
   image: lmsysorg/sglang:v0.5.11-cu130
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6e1249adfc..4d30e92b22 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4581,3 +4581,12 @@
     - "Pin the EAGLE3 drafter to TRITON_ATTN via the speculative-config attention_backend override; without it the drafter falls back to a slow default backend."
     - "Speeds up the draft forward measured on MI355X TP4 MXFP8, 8k/1k throughput gains growing with concurrency vs no override."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2107
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang
+  description:
+    - "Add DeepSeek-V4-Pro NVFP4 GB300 1k1k STP disagg scenarios to the existing dsv4-fp4-gb300-dynamo-sglang key (7 fixed-seq-len entries alongside the existing 8k1k scenarios)."
+    - "Topologies: 1P1D DEP8 and DEP16 (custom conc sweep), 1P4D and 1P6D TP4 (sa-bench conc sweep), 2P1D DEP12/DEP16/DEP48 (custom conc sweep)."
+    - "Recipes live under benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/1k1k/, container aligned with the existing 8k1k image (lmsysorg/sglang:nightly-dev-cu13-20260520-425dffbd)."
+    - "launch_gb300-nv.sh gains a dynamo-sglang + dsv4 overlay branch (srt-slurm sa-submission-q2-2026 base, in-tree recipes overlaid on top)."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2099
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index aaf48134fa..4731a47adc 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -156,6 +156,12 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     git checkout aflowers/gb200-dsv4-recipes
     mkdir -p recipes/vllm/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    git checkout sa-submission-q2-2026
+    mkdir -p recipes/sglang/deepseek-v4
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"