From ba76967e9ac34ae0a44819bb0e0df9070452e01d Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 27 Jul 2026 15:17:01 -0700
Subject: [PATCH 1/6] feat: add DSV4 FP4 B300 Dynamo-SGLang MTP config
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：新增 DSV4 FP4 B300 Dynamo-SGLang MTP 配置，并接入启用聊天模板输入和 EAGLE 投机解码的共享 srt-slurm 配方矩阵。
---
 configs/nvidia-master.yaml | 142 +++++++++++++++++++++++++++++++++++++
 perf-changelog.yaml        |   8 +++
 runners/launch_b300-nv.sh  |  25 +++++--
 3 files changed, 171 insertions(+), 4 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 36ae816809..a404e94837 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6410,6 +6410,148 @@ dsv4-fp4-gb300-dynamo-sglang-mtp:
           ep: 8
           dp-attn: true
 
+# DSV4 NVFP4 B300 disaggregated dynamo-sglang MTP configuration.
+# Each search-space entry selects one variant from the shared srt-slurm recipe
+# matrix. Runs on the b300 pool via runners/launch_b300-nv.sh. Prefill workers
+# are TP4/DEP4 (two workers per 8-GPU B300 node); decode is DEP8 (one full node)
+# except the TP4 topologies. EAGLE draft depth is tapered down as concurrency
+# grows (3 steps / 4 draft tokens -> 2/3 at 6p1d -> 1/2 at 8p1d).
+dsv4-fp4-b300-dynamo-sglang-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: dynamo-sglang
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1p1d-tp4-tp4 (pure TP, no dp-attn). 2 nodes (1P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [1]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:override_tp4_tp4"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      # Scale-out: 1x DEP4 prefill + 6x TP4 decode. 4 nodes (1P + 3D).
+      - spec-decoding: "mtp"
+        conc-list: [8, 32, 64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:override_dep4_tp4"
+        decode:
+          num-worker: 6
+          tp: 4
+          ep: 1
+          dp-attn: false
+      # 1p1d-dep4-dep8: 1x DEP4 prefill + DEP8 decode. 2 nodes (1P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[0]"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 1p2d-dep4-dep8: 1x DEP4 prefill + 2x DEP8 decode. 3 nodes (1P + 2D).
+      - spec-decoding: "mtp"
+        conc-list: [256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[1]"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 2p1d-dep4-dep8: 2x DEP4 prefill + DEP8 decode. 2 nodes (1P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [512]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[2]"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 4p1d-dep4-dep8: 4x DEP4 prefill + DEP8 decode. 3 nodes (2P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [1024]
+        prefill:
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[3]"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 6p1d-dep4-dep8: 6x DEP4 prefill + DEP8 decode. 4 nodes (3P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [2048]
+        prefill:
+          num-worker: 6
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[4]"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 8p1d-dep4-dep8. 5 nodes (4P + 1D).
+      - spec-decoding: "mtp"
+        conc-list: [3072]
+        prefill:
+          num-worker: 8
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[5]"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 kimik2.5-int4-h100-vllm:
   image: vllm/vllm-openai:v0.22.0
   model: moonshotai/Kimi-K2.5
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0524d0c0c9..1b6bb2212a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5116,3 +5116,11 @@
   description:
     - "Bump image from lmsysorg/sglang:v0.5.14-rocm720-mi35x to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260726"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2349
+
+- config-keys:
+    - dsv4-fp4-b300-dynamo-sglang-mtp
+  description:
+    - "Add DSV4 NVFP4 B300 disaggregated Dynamo SGLang MTP configuration for 8k/1k"
+    - "Use the shared srt-slurm recipe matrix from NVIDIA/srt-slurm#281 with EAGLE speculative decoding and chat-template inputs"
+    - "Image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 098e984be5..f4379afcd9 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -41,6 +41,9 @@ elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-
     fi
     export MODEL_PATH="${SELECTED_MODEL_PATH:-/data/models/dsv4-pro}"
     export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
+elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-sglang" ]]; then
+    export MODEL_PATH="${MODEL_PATH:-/scratch/models/DeepSeek-V4-Pro}"
+    export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
 elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-vllm" ]]; then
     export MODEL_PATH="/data/models/MiniMax-M2.5-NVFP4"
     export SRT_SLURM_MODEL_PREFIX="minimax-m2.5-nvfp4"
@@ -54,7 +57,7 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" && $FRAMEWORK == "dy
     export MODEL_PATH="/data/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm or dynamo-sglang, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm"
     exit 1
 fi
 
@@ -76,6 +79,9 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     git checkout aflowers/vllm-gb200-v0.20.0
     mkdir -p recipes/vllm/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
+    git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR" || exit 1
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
@@ -184,10 +190,18 @@ if [[ -z "$CONFIG_FILE" ]]; then
     exit 1
 fi
 
-# Override the job name in the config file with the runner name
-sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
+# Resolve the recipe path before editing it. CONFIG_FILE may include an
+# srt-slurm matrix selector such as :zip_override_dep4_dep8[0].
+CONFIG_PATH="${CONFIG_FILE%%:*}"
+if [[ ! -f "$CONFIG_PATH" ]]; then
+    echo "Error: CONFIG_FILE does not exist after srt-slurm setup: $CONFIG_PATH" >&2
+    exit 1
+fi
+
+# Override the job name in the recipe with the runner name.
+sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_PATH"
 if [[ "$MODEL_PREFIX" == "minimaxm3" && -n "$MINIMAX_M3_SLURM_EXCLUDED_NODELIST" ]]; then
-    sed -i "/^name:.*/a sbatch_directives:\n  exclude: \"${MINIMAX_M3_SLURM_EXCLUDED_NODELIST}\"" "$CONFIG_FILE"
+    sed -i "/^name:.*/a sbatch_directives:\n  exclude: \"${MINIMAX_M3_SLURM_EXCLUDED_NODELIST}\"" "$CONFIG_PATH"
 fi
 SRTCTL_APPLY_ARGS=(
     -f "$CONFIG_FILE"
@@ -201,6 +215,9 @@ SRTCTL_APPLY_ARGS=(
 if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && ( "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml || "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/*-tp1-*.yaml ) ]]; then
     SRTCTL_APPLY_ARGS+=(--no-preflight)
 fi
+if [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" && "$MODEL_PATH" == /scratch/models/* ]]; then
+    SRTCTL_APPLY_ARGS+=(--no-preflight)
+fi
 if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
     SRTCTL_APPLY_ARGS+=(--setup-script "$SRTCTL_SETUP_SCRIPT")
 fi

From 51c3739c2fbf0755ab66d4dc87745bab319a0747 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 27 Jul 2026 15:18:29 -0700
Subject: [PATCH 2/6] chore: link perf changelog to #2363
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将性能变更日志链接更新为 SemiAnalysisAI/InferenceX#2363。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1b6bb2212a..6c9fbabea3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5123,4 +5123,4 @@
     - "Add DSV4 NVFP4 B300 disaggregated Dynamo SGLang MTP configuration for 8k/1k"
     - "Use the shared srt-slurm recipe matrix from NVIDIA/srt-slurm#281 with EAGLE speculative decoding and chat-template inputs"
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2363

From 25eddd2c2e43dcde355b50ad518db758830dfc0b Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 27 Jul 2026 15:33:12 -0700
Subject: [PATCH 3/6] fix: use B300 DSV4 NVFP4 model path
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 B300 DSV4 Dynamo-SGLang 配置指向服务器上预置的 DeepSeek-V4-Pro-NVFP4 模型路径。
---
 runners/launch_b300-nv.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index f4379afcd9..904aa8f059 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -42,7 +42,7 @@ elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-
     export MODEL_PATH="${SELECTED_MODEL_PATH:-/data/models/dsv4-pro}"
     export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
 elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-sglang" ]]; then
-    export MODEL_PATH="${MODEL_PATH:-/scratch/models/DeepSeek-V4-Pro}"
+    export MODEL_PATH="${MODEL_PATH:-/scratch/models/DeepSeek-V4-Pro-NVFP4}"
     export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
 elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-vllm" ]]; then
     export MODEL_PATH="/data/models/MiniMax-M2.5-NVFP4"

From d685ccfb4bb9b0fec711d8e8890de29851c62ce3 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 27 Jul 2026 16:12:38 -0700
Subject: [PATCH 4/6] fix: pin srt-slurm recipe revision
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

固定 srt-slurm 配方版本，确保 DSV4 B300 Dynamo-SGLang 运行使用可复现的配置。
---
 runners/launch_b300-nv.sh | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 904aa8f059..b76b9a9609 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -82,6 +82,8 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
+    git fetch origin 2065f6e7e798b48c0dc8d20a33c1f75427659dec || exit 1
+    git checkout 2065f6e7e798b48c0dc8d20a33c1f75427659dec || exit 1
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1

From 2fda5e974357c752c947cca5842f584c3c9848bf Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Tue, 28 Jul 2026 01:02:46 -0700
Subject: [PATCH 5/6] fix(dsv4): pin UCX-enabled srt-slurm revision
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

修复（dsv4）：固定包含 UCX CUDA 传输配置的 srt-slurm 版本。
---
 runners/launch_b300-nv.sh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index b76b9a9609..43fa9c43b3 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -82,8 +82,8 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
-    git fetch origin 2065f6e7e798b48c0dc8d20a33c1f75427659dec || exit 1
-    git checkout 2065f6e7e798b48c0dc8d20a33c1f75427659dec || exit 1
+    git fetch origin cb94d592dfd03a7f3f1ede9f0a3bbbb8a8452ab7 || exit 1
+    git checkout cb94d592dfd03a7f3f1ede9f0a3bbbb8a8452ab7 || exit 1
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1

From 6b1d6f15499c032be91bf47287c943d1ef05b47c Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Wed, 29 Jul 2026 00:09:08 -0700
Subject: [PATCH 6/6] fix: use checked-in DSV4 MTP recipes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Pin NVIDIA/srt-slurm main, overlay the checked-in recipes, and enable CUDA-aware UCX transport for every MTP topology.

修复：固定 NVIDIA/srt-slurm main 提交，覆盖仓库内置配方，并为所有 MTP 拓扑启用支持 CUDA 的 UCX 传输。
---
 .../8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml  | 134 +++++++++++++++++
 .../8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml    | 125 ++++++++++++++++
 .../8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml  | 135 ++++++++++++++++++
 .../8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml   | 129 +++++++++++++++++
 .../8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml  | 135 ++++++++++++++++++
 .../8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml  | 135 ++++++++++++++++++
 .../8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml  | 135 ++++++++++++++++++
 .../8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml  | 135 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  29 ++--
 perf-changelog.yaml                           |   4 +-
 runners/launch_b300-nv.sh                     |   5 +-
 11 files changed, 1083 insertions(+), 18 deletions(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..a07d5bb0a7
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
@@ -0,0 +1,134 @@
+name: "disagg-b300-1p1d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "2048"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.94
+      max-running-requests: 3072
+      cuda-graph-max-bs: 256
+      swa-full-tokens-ratio: 0.15
+      context-length: 16384
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "256"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
new file mode 100644
index 0000000000..75c1815c8a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
@@ -0,0 +1,125 @@
+name: "disagg-b300-1p1d-tp4-tp4-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 4
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 8
+      cuda-graph-max-bs: 8
+      chunked-prefill-size: 32768
+
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.9
+      max-running-requests: 8
+      cuda-graph-max-bs: 8
+      swa-full-tokens-ratio: 0.1
+      context-length: 16384
+
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "1"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..fef5285ceb
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
@@ -0,0 +1,135 @@
+name: "disagg-b300-1p2d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 2
+  decode_workers: 2
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "2048"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 256
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.94
+      max-running-requests: 3072
+      cuda-graph-max-bs: 256
+      swa-full-tokens-ratio: 0.15
+      context-length: 16384
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "256"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
new file mode 100644
index 0000000000..d47b469945
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
@@ -0,0 +1,129 @@
+name: "disagg-b300-1p6d-dep4-tp4-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_nodes: 3
+  decode_workers: 6
+  gpus_per_decode: 4
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 128
+      cuda-graph-max-bs: 128
+      chunked-prefill-size: 32768
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.9
+      max-running-requests: 128
+      cuda-graph-max-bs: 128
+      swa-full-tokens-ratio: 0.1
+      context-length: 16384
+
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "8x32x64"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..e796e2c680
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
@@ -0,0 +1,135 @@
+name: "disagg-b300-2p1d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "4096"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 512
+      cuda-graph-max-bs: 512
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.9
+      max-running-requests: 3072
+      cuda-graph-max-bs: 512
+      swa-full-tokens-ratio: 0.15
+      context-length: 16384
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "512"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..681f350d90
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
@@ -0,0 +1,135 @@
+name: "disagg-b300-4p1d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 2
+  prefill_workers: 4
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "4096"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 1024
+      cuda-graph-max-bs: 1024
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+      mem-fraction-static: 0.9
+      max-running-requests: 3072
+      cuda-graph-max-bs: 1024
+      swa-full-tokens-ratio: 0.15
+      context-length: 16384
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "1024"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..4813ac793f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
@@ -0,0 +1,135 @@
+name: "disagg-b300-6p1d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 3
+  prefill_workers: 6
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "4096"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 1024
+      cuda-graph-max-bs: 1024
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 2
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 3
+
+      mem-fraction-static: 0.9
+      max-running-requests: 4096
+      cuda-graph-max-bs: 1024
+      swa-full-tokens-ratio: 0.15
+      context-length: 9216
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "2048"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml
new file mode 100644
index 0000000000..ed3b27e906
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml
@@ -0,0 +1,135 @@
+name: "disagg-b300-8p1d-dep4-dep8-mtp"
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 8
+
+dynamo:
+  hash: "41882ae9b07232eed4850fb1daf8c958abb2556a"
+  install: true
+
+model:
+  path: "deepseek-v4-pro"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
+  precision: "fp4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 4
+  prefill_workers: 8
+  gpus_per_prefill: 4
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 8
+
+backend:
+  type: sglang
+
+  prefill_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "9216"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    PYTHONUNBUFFERED: "1"
+    SGLANG_RADIX_FORCE_MISS: "1"
+    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: "max"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "4096"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1"
+    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    MC_FORCE_MNNVL: "1"
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: "100000"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "100000"
+    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: "1"
+    UCX_TLS: "cuda_copy,rc"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "0"
+
+  sglang_config:
+    prefill:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "prefill"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      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: 0.9
+      max-running-requests: 1024
+      cuda-graph-max-bs: 1024
+      chunked-prefill-size: 32768
+      stream-interval: 60
+
+    decode:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      model-path: "/model/"
+      trust-remote-code: true
+      tool-call-parser: deepseekv4
+
+      disaggregation-mode: "decode"
+      watchdog-timeout: 86400
+      disaggregation-transfer-backend: nixl
+
+      tensor-parallel-size: 8
+      data-parallel-size: 8
+      expert-parallel-size: 8
+
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-a2a-backend: "megamoe"
+
+      speculative-algo: "EAGLE"
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 2
+
+      mem-fraction-static: 0.85
+      max-running-requests: 8192
+      cuda-graph-max-bs: 1280
+      swa-full-tokens-ratio: 0.1
+      context-length: 9216
+      stream-interval: 60
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  random_range_ratio: 0.8
+  concurrencies: "3072"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 47fb3918a8..d4becea465 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6441,12 +6441,13 @@ dsv4-fp4-gb300-dynamo-sglang-mtp:
           ep: 8
           dp-attn: true
 
-# DSV4 FP4 B300 disaggregated dynamo-sglang MTP configuration.
-# Each search-space entry selects one variant from the shared srt-slurm recipe
-# matrix. Runs on the b300 pool via runners/launch_b300-nv.sh. Prefill workers
-# are TP4/DEP4 (two workers per 8-GPU B300 node); decode is DEP8 (one full node)
-# except the TP4 topologies. EAGLE draft depth is tapered down as concurrency
-# grows (3 steps / 4 draft tokens -> 2/3 at 6p1d -> 1/2 at 8p1d).
+# DSV4 B300 disaggregated dynamo-sglang MTP configuration.
+# Each search-space entry maps one-to-one to a checked-in srt-slurm recipe under
+# recipes/sglang/deepseek-v4/8k1k/. Runs on the b300 pool via
+# runners/launch_b300-nv.sh. Prefill workers are TP4/DEP4 (two workers per
+# 8-GPU B300 node); decode is DEP8 (one full node) except the TP4 topologies.
+# EAGLE draft depth is tapered down as concurrency grows
+# (3 steps / 4 draft tokens -> 2/3 at 6p1d -> 1/2 at 8p1d).
 dsv4-fp4-b300-dynamo-sglang-mtp:
   image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4
   model: deepseek-ai/DeepSeek-V4-Pro
@@ -6471,7 +6472,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:override_tp4_tp4"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml"
         decode:
           num-worker: 1
           tp: 4
@@ -6486,7 +6487,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:override_dep4_tp4"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml"
         decode:
           num-worker: 6
           tp: 4
@@ -6501,7 +6502,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[0]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
@@ -6516,7 +6517,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[1]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 2
           tp: 8
@@ -6531,7 +6532,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[2]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
@@ -6546,7 +6547,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[3]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
@@ -6561,7 +6562,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[4]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
@@ -6576,7 +6577,7 @@ dsv4-fp4-b300-dynamo-sglang-mtp:
           ep: 4
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4-pro/sglang/b300-fp4/8k1k-mtp.yaml:zip_override_dep4_dep8[5]"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml"
         decode:
           num-worker: 1
           tp: 8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 055317cc19..4d7d9b646c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5163,7 +5163,7 @@
 - config-keys:
     - dsv4-fp4-b300-dynamo-sglang-mtp
   description:
-    - "Add DSV4 FP4 B300 disaggregated Dynamo SGLang MTP configuration for 8k/1k"
-    - "Use the shared srt-slurm recipe matrix from NVIDIA/srt-slurm#281 with EAGLE speculative decoding and chat-template inputs"
+    - "Add DSV4 B300 disaggregated Dynamo SGLang MTP configuration for 8k/1k"
+    - "Add 8 checked-in srt-slurm recipes with EAGLE speculative decoding and chat-template inputs"
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2363
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 2c6a07d20b..f647446f02 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -82,8 +82,9 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
-    git fetch origin cb94d592dfd03a7f3f1ede9f0a3bbbb8a8452ab7 || exit 1
-    git checkout cb94d592dfd03a7f3f1ede9f0a3bbbb8a8452ab7 || exit 1
+    git checkout c180328b98c3793ca84a1e24a030f90545eb7d5d || exit 1
+    mkdir -p recipes/sglang/deepseek-v4
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml ]]; then
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1