From ac2327418b72fa90d097fbaa41a9c1586292d562 Mon Sep 17 00:00:00 2001
From: jasonlizhengjian <jasonlizhengjian@gmail.com>
Date: Tue, 2 Jun 2026 11:14:48 -0700
Subject: [PATCH 1/3] Add MiniMax-M2.5 FP4 GB300 Dynamo vLLM recipes

---
 .github/configs/nvidia-master.yaml            | 174 ++++++++++++++++++
 .../minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml    |  73 ++++++++
 .../vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml     |  73 ++++++++
 .../vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml     |  70 +++++++
 .../vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml      |  72 ++++++++
 .../vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml      |  68 +++++++
 .../vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml    |  70 +++++++
 .../vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml    |  68 +++++++
 .../vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml     |  70 +++++++
 .../vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml     |  70 +++++++
 .../vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml      |  68 +++++++
 .../vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml    |  68 +++++++
 .../vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml    |  68 +++++++
 perf-changelog.yaml                           |   7 +
 runners/launch_gb300-nv.sh                    |  11 +-
 15 files changed, 1029 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index b8db07e0d..69f56fa08 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -9261,6 +9261,180 @@ qwen3.5-fp8-h100-sglang-mtp:
       search-space:
       - { tp: 8, ep: 8, conc-start: 4, conc-end: 32, spec-decoding: mtp }
 
+minimaxm2.5-fp4-gb300-dynamo-vllm:
+  image: vllm/vllm-openai:v0.20.1
+  model: nvidia/MiniMax-M2.5-NVFP4
+  model-prefix: minimaxm2.5
+  runner: gb300
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - conc-list: [2, 4, 16]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [4, 8, 16, 64]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [32, 64, 128]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+      - conc-list: [64, 128, 256, 512, 1024]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml"
+        decode:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: false
+      - conc-list: [2048]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml"
+        decode:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+      - conc-list: [6144, 8192]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml"
+        decode:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+      - conc-list: [1024, 2048, 4096]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - conc-list: [2, 4, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [32, 64, 128, 256]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+      - conc-list: [64, 128]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+      - conc-list: [256]
+        prefill:
+          num-worker: 4
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+      - conc-list: [1024, 2048]
+        prefill:
+          num-worker: 4
+          tp: 1
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 glm5-fp4-gb300-dynamo-sglang:
   image: lmsysorg/sglang:v0.5.11-cu130
   model: nvidia/GLM-5-NVFP4
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml
new file mode 100644
index 000000000..c7f7e28af
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d-c6144.yaml
@@ -0,0 +1,73 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-2p3xdep2-c6144"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 3
+  prefill_workers: 2
+  decode_workers: 3
+  gpus_per_prefill: 1
+  gpus_per_decode: 2
+  spread_workers: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 128
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      max-num-seqs: 864
+      gpu-memory-utilization: 0.90
+      stream-interval: 128
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "6144x8192"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml
new file mode 100644
index 000000000..adaf6f271
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep2-2p3d.yaml
@@ -0,0 +1,73 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-2p3xdep2"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 3
+  prefill_workers: 2
+  decode_workers: 3
+  gpus_per_prefill: 1
+  gpus_per_decode: 2
+  spread_workers: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      data-parallel-size: 2
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      max-num-seqs: 864
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "2048"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml
new file mode 100644
index 000000000..28427e002
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/dep8-2p1d.yaml
@@ -0,0 +1,70 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-2p1xdep8"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 128
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      max-num-seqs: 864
+      gpu-memory-utilization: 0.90
+      stream-interval: 128
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "1024x2048x4096"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml
new file mode 100644
index 000000000..eee93c9f8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p1d.yaml
@@ -0,0 +1,72 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-focus-tp4-1p1d"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+    UCX_RCACHE_MAX_UNRELEASED: "1024"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+    UCX_RCACHE_MAX_UNRELEASED: "1024"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: false
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "2x4x16"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml
new file mode 100644
index 000000000..10ba980ca
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4-1p2d.yaml
@@ -0,0 +1,68 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-focus-tp4-1p2d"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: false
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "4x8x16x64"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml
new file mode 100644
index 000000000..ebff26fb0
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p1d.yaml
@@ -0,0 +1,70 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-focus-tp4ep-1p1d"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    UCX_TLS: "cuda_copy,rc"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "32x64x128"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml
new file mode 100644
index 000000000..5353e4dd0
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/1k1k/tp4ep-1p3d.yaml
@@ -0,0 +1,68 @@
+name: "minimax-m2.5-vllm-disagg-gb300-decode-focus-tp4ep-1p3d"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 3
+  prefill_workers: 1
+  decode_workers: 3
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 2048
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "64x128x256x512x1024"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml
new file mode 100644
index 000000000..d3c777618
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep4-4p1d.yaml
@@ -0,0 +1,70 @@
+name: "minimax-m2.5-vllm-disagg-gb300-8k1k-4p1xdep4"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 4
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      max-num-seqs: 864
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "256"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml
new file mode 100644
index 000000000..a56c095af
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/dep8-4p1d.yaml
@@ -0,0 +1,70 @@
+name: "minimax-m2.5-vllm-disagg-gb300-8k1k-4p1xdep8"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 4
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 8
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      max-num-seqs: 864
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1024x2048"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml
new file mode 100644
index 000000000..a92975c57
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4-1p1d.yaml
@@ -0,0 +1,68 @@
+name: "minimax-m2.5-vllm-disagg-gb300-8k1k-1p1xtp4"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: false
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "2x4x8x16"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml
new file mode 100644
index 000000000..53daeafbd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-1p1d.yaml
@@ -0,0 +1,68 @@
+name: "minimax-m2.5-vllm-disagg-gb300-8k1k-1p1xtp4ep"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "32x64x128x256"
+  random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml
new file mode 100644
index 000000000..163d412f5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/8k1k/tp4ep-2p1d.yaml
@@ -0,0 +1,68 @@
+name: "minimax-m2.5-vllm-disagg-gb300-8k1k-2p1xtp4ep"
+
+model:
+  path: "minimax-m2.5-nvfp4"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260526"
+
+setup_script: install-deps.sh
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 1
+  gpus_per_decode: 4
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+      stream-interval: 32
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      no-enable-prefix-caching: true
+      max-model-len: 9280
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+      gpu-memory-utilization: 0.90
+      stream-interval: 32
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "64x128"
+  random_range_ratio: 0.8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 28523da86..c4f5232b9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3409,3 +3409,10 @@
   description:
     - "Add MiniMax-M2.5 NVFP4 GB200 disaggregated multinode vLLM benchmarks via Dynamo"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1642
+
+- config-keys:
+    - minimaxm2.5-fp4-gb300-dynamo-vllm
+  description:
+    - "Add MiniMax-M2.5 NVFP4 GB300 disaggregated multinode vLLM benchmarks via Dynamo"
+    - "Add 1k1k/8k1k minimax recipe set under benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5/"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1641
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index b47e103fd..0a24120f1 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -45,8 +45,11 @@ elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp4" ]]; then
 elif [[ $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH=/scratch/models/GLM-5-FP8
     export SRT_SLURM_MODEL_PREFIX="glm-5-fp8"
+elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp4" ]]; then
+    export MODEL_PATH=/data/models/MiniMax-M2.5-NVFP4
+    export SRT_SLURM_MODEL_PREFIX="minimax-m2.5-nvfp4"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4, glm5-fp8"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4, glm5-fp8, minimaxm2.5-fp4"
     exit 1
 fi
 
@@ -137,6 +140,12 @@ elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" ]]; then
     git checkout sa-submission-q2-2026
     mkdir -p recipes/sglang/glm5
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5" recipes/sglang/glm5
+elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm2.5" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    git checkout main
+    mkdir -p recipes/vllm/minimax-m2.5
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m2.5" recipes/vllm/minimax-m2.5
 else
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"

From 97fbd88df1da6b8577d8b4db553e727738b0d733 Mon Sep 17 00:00:00 2001
From: jasonlizhengjian <jasonlizhengjian@gmail.com>
Date: Tue, 2 Jun 2026 11:31:05 -0700
Subject: [PATCH 2/3] fix: pin minimax gb300 sweep to nv runners

---
 .github/configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 69f56fa08..3a4f1a4f1 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -9265,7 +9265,7 @@ minimaxm2.5-fp4-gb300-dynamo-vllm:
   image: vllm/vllm-openai:v0.20.1
   model: nvidia/MiniMax-M2.5-NVFP4
   model-prefix: minimaxm2.5
-  runner: gb300
+  runner: gb300-nv
   precision: fp4
   framework: dynamo-vllm
   multinode: true

From 45759a70815cb4ce7a43223e094bb13e4d99e4e5 Mon Sep 17 00:00:00 2001
From: jasonlizhengjian <jasonlizhengjian@gmail.com>
Date: Tue, 2 Jun 2026 18:24:42 -0700
Subject: [PATCH 3/3] Fix GB300 eval artifact copy

---
 runners/launch_gb300-nv.sh | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 0a24120f1..4677e5004 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -379,7 +379,9 @@ if [[ "${RUN_EVAL:-false}" == "true" || "${EVAL_ONLY:-false}" == "true" ]]; then
         shopt -s nullglob
         for eval_file in "$EVAL_DIR"/*; do
             [ -f "$eval_file" ] || continue
-            cp "$eval_file" "$GITHUB_WORKSPACE/"
+            eval_dest="$GITHUB_WORKSPACE/$(basename "$eval_file")"
+            rm -f "$eval_dest"
+            cp "$eval_file" "$eval_dest"
             echo "Copied eval artifact: $(basename "$eval_file")"
         done
         shopt -u nullglob