diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 46d26b329a..fb7ef082c2 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -12308,6 +12308,191 @@ qwen3.5-fp4-b200-trt:
       - { tp: 4, ep: 4, dp-attn: true, conc-list: [1024] }
       - { tp: 8, ep: 8, dp-attn: true, conc-list: [256, 512, 1024] }
 
+# MiniMax-M3 GB200 disagg sweep — adapted from NV B300 PR #1863.
+# All prefill DEP4 (TP1 DP4 EP, 4 GPU/worker). Decode: Marlin, TEP8, DEP8, TEP4.
+# 4 GPU/node (GB200 NVL72). FLASHINFER attention with FP8 KV cache.
+minimaxm3-fp8-gb200-dynamo-vllm:
+  image: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: gb200
+  precision: fp8
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      # 1p1d DEP4+DEP8, 3n: conc 1024,4096
+      - conc-list: [1024, 4096]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 1p1d DEP4+TEP8, 3n: conc 128,256
+      - conc-list: [128, 256]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 1p2d DEP4+DEP8, 5n: conc 1024,4096
+      - conc-list: [1024, 4096]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 1p2d DEP4+TEP8, 5n: conc 4,16,64
+      - conc-list: [4, 16, 64]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 1p4d DEP4+TP4 Marlin, 5n: conc 32
+      - conc-list: [32]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml"
+        decode:
+          num-worker: 4
+          tp: 4
+          ep: 1
+          dp-attn: false
+
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1p2d DEP4+DEP8, 5n: conc 512
+      - conc-list: [512]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 1p2d DEP4+TEP4, 3n: conc 4
+      - conc-list: [4]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: false
+
+      # 1p2d DEP4+TEP8, 5n: conc 4,16,32,64,128
+      - conc-list: [4, 16, 32, 64, 128]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: false
+
+      # 2p2d DEP4+DEP8, 6n: conc 512,1024
+      - conc-list: [512, 1024]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 3p2d DEP4+DEP8, 7n: conc 4096
+      - conc-list: [4096]
+        prefill:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+      # 5p2d DEP4+DEP8, 9n: conc 4096
+      - conc-list: [4096]
+        prefill:
+          num-worker: 5
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml"
+        decode:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 qwen3.5-fp4-b200-trt-mtp:
   image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc18
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/minimax-m3-gb200-vllm-fixes.sh b/benchmarks/multi_node/srt-slurm-recipes/configs/minimax-m3-gb200-vllm-fixes.sh
new file mode 100755
index 0000000000..c0eed0a517
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/minimax-m3-gb200-vllm-fixes.sh
@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3 - <<'PYEOF'
+from importlib.util import find_spec
+from pathlib import Path
+
+spec = find_spec("vllm")
+if not spec or not spec.origin:
+    raise RuntimeError("vllm is not installed")
+root = Path(spec.origin).parent
+patches = {
+    root / "distributed/device_communicators/flashinfer_all_reduce.py": [
+        (
+            "            comm_backend=comm_backend,\n"
+            "            group=group,\n",
+            "            comm_backend=comm_backend,\n"
+            '            force_oneshot_support=backend == "mnnvl",\n'
+            "            group=group,\n",
+        ),
+    ],
+    root / "models/minimax_m3/nvidia/sparse_attention_msa.py": [
+        (
+            "            prefill_topk = topk[:, nd:num_tokens, :]\n",
+            "            prefill_topk = topk[:, nd:num_tokens, :].contiguous()\n",
+        ),
+    ],
+}
+for path, edits in patches.items():
+    source = path.read_text()
+    for old, new in edits:
+        if new in source:
+            continue
+        if source.count(old) != 1:
+            raise RuntimeError(f"missing or ambiguous patch anchor in {path}")
+        source = source.replace(old, new, 1)
+    path.write_text(source)
+PYEOF
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml
new file mode 100644
index 0000000000..710c4d9796
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-1p1d-dep4-dep8-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 1D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 2 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 4096
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "1024x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml
new file mode 100644
index 0000000000..c1b9cf32d6
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml
@@ -0,0 +1,105 @@
+name: "minimax-m3-vllm-disagg-gb200-1p1d-dep4-tep8-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 1D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 2 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 8
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-num-seqs: 4096
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 8196
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "128x256"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml
new file mode 100644
index 0000000000..907633ba74
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-1p2d-dep4-dep8-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 4096
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "1024x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml
new file mode 100644
index 0000000000..2e2bbc12a9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml
@@ -0,0 +1,105 @@
+name: "minimax-m3-vllm-disagg-gb200-1p2d-dep4-tep8-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 8
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-num-seqs: 4096
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 8196
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "4x16x64"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml
new file mode 100644
index 0000000000..429aa015ea
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-1p4d-dep4-tp4-marlin-fp8-1k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 4D TP4 decode (TP4, 4 GPU/worker = 1 node each).
+# Decode uses Marlin MoE backend with expert-parallel disabled (1p4d only).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 4
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 2048
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 4
+      moe-backend: marlin
+      enable-expert-parallel: false
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 2304
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 8196
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "32"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml
new file mode 100644
index 0000000000..4129408ebd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-1p2d-dep4-dep8-fp8-8k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "512"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml
new file mode 100644
index 0000000000..5bde9ac9be
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml
@@ -0,0 +1,105 @@
+name: "minimax-m3-vllm-disagg-gb200-1p2d-dep4-tep4-fp8-8k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D TEP4 decode (TP4 EP, 4 GPU/worker = 1 node each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 2 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 4
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-num-seqs: 512
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 4096
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml
new file mode 100644
index 0000000000..56ed02e5e8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml
@@ -0,0 +1,105 @@
+name: "minimax-m3-vllm-disagg-gb200-1p2d-dep4-tep8-fp8-8k1k"
+
+# 1P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 8
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 32
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 4096
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4x16x32x64x128"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml
new file mode 100644
index 0000000000..22ebfc7836
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-2p2d-dep4-dep8-fp8-8k1k"
+
+# 2P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 1 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 4
+  prefill_workers: 2
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "512x1024"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml
new file mode 100644
index 0000000000..ba1db206e4
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-3p2d-dep4-dep8-fp8-8k1k"
+
+# 3P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 2 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 3
+  decode_nodes: 4
+  prefill_workers: 3
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml
new file mode 100644
index 0000000000..a54ddcb226
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb200-5p2d-dep4-dep8-fp8-8k1k"
+
+# 5P DEP4 prefill (TP1 DP4 EP, 4 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB200 has 4 GPUs/node.
+# Nodes: 3 prefill + 4 decode (+ head/infra).
+
+model:
+  path: "minimax-m3-mxfp8"
+  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+  precision: "fp8"
+
+
+dynamo:
+  install: true
+  version: 1.3.0.dev20260614
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+
+srun_options:
+  mem: "0"
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  prefill_nodes: 5
+  decode_nodes: 4
+  prefill_workers: 5
+  decode_workers: 2
+  gpus_per_prefill: 4
+  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"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_FLOAT32_MATMUL_PRECISION: "high"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
+    UCX_CUDA_IPC_ENABLE_MNNVL: "y"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 4
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 16384
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}'
+      tensor-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 128
+      gpu-memory-utilization: 0.90
+      max-model-len: 9472
+      language-model-only: true
+      kv-cache-dtype: fp8
+      stream-interval: 100
+      max-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096"
+  req_rate: "inf"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 1cbadb492f..0028a654e6 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4132,6 +4132,14 @@
     - "Backport vllm-project/vllm#45879 so NIXL validates heterogeneous-TP KV block lengths using the GQA KV-head ratio."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1893
 
+- config-keys:
+    - minimaxm3-fp8-gb200-dynamo-vllm
+  description:
+    - "Initial addition of MiniMax-M3 FP8 GB200 disaggregated Dynamo-vLLM benchmark using vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
+    - "Allocate FlashInfer MNNVL workspace for one-shot all-reduce and materialize the MSA prefill top-k slice before CSR construction"
+    - "Preserve current Qwen3.5 and Kimi-K2.5 GB200 launcher paths while adding MiniMax-M3 shared-FS staging and atomic image import"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1938
+
 - config-keys:
     - minimaxm3-fp8-gb300-dynamo-vllm
   description:
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 5c018d576d..065ff7da70 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -60,8 +60,11 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
     elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp8" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/MiniMax-M2.5"
         export SRT_SLURM_MODEL_PREFIX="minimax-m2.5-fp8"
+    elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+        export MODEL_PATH="/mnt/lustre01/models/MiniMax-M3-MXFP8"
+        export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
     else
-        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8"
+        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
         exit 1
     fi
 else
@@ -74,15 +77,22 @@ export SLURM_ACCOUNT="benchmark"
 
 NGINX_IMAGE="nginx:1.27.4"
 
-# === Cluster diagnostic probe (minimax only) ===
+uses_watchtower_shared_fs() {
+    case "$MODEL_PREFIX" in
+        minimaxm2.5|minimaxm3|kimik2.5) return 0 ;;
+        *) return 1 ;;
+    esac
+}
+
+# === Cluster diagnostic probe for watchtower-hosted sweeps ===
 # The gb200-nv_* runners may be hosted on different physical clusters
 # (e.g., the legacy NVIDIA Lustre cluster vs Oracle Cloud "watchtower").
 # Print enough info to identify the layout, then pick a writable
 # squash dir on a path that's also visible to compute nodes. Falls
 # back to the legacy sa-shared path so other configs are untouched.
 SQUASH_DIR="/mnt/lustre01/users-public/sa-shared"
-if [[ $MODEL_PREFIX == "minimaxm2.5" || $MODEL_PREFIX == "kimik2.5" ]]; then
-    echo "=== cluster diagnostic (minimax sweep) ==="
+if uses_watchtower_shared_fs; then
+    echo "=== cluster diagnostic (watchtower sweep) ==="
     echo "USER=$(id -un) UID=$(id -u) GID=$(id -g) GROUPS=$(id -Gn)"
     echo "HOME=$HOME"
     echo "HOSTNAME=$(hostname -f 2>/dev/null || hostname)"
@@ -128,8 +138,27 @@ fi
 SQUASH_FILE="${SQUASH_DIR}/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 NGINX_SQUASH_FILE="${SQUASH_DIR}/$(echo "$NGINX_IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 
-enroot import -o $SQUASH_FILE docker://$IMAGE
-enroot import -o $NGINX_SQUASH_FILE docker://$NGINX_IMAGE
+# Concurrent matrix jobs import to the same shared-FS squash path.
+# Serialize imports and atomically replace invalid images so readers never
+# observe a partially written squash file.
+import_squash() {
+    local squash="$1" image="$2"
+    local lock="${squash}.lock"
+    (
+        exec 9>"$lock"
+        flock -w 1800 9 || { echo "Failed to acquire lock for $squash" >&2; exit 1; }
+        if unsquashfs -l "$squash" > /dev/null 2>&1; then
+            echo "Squash file already exists and is valid, skipping import: $squash"
+        else
+            rm -f "$squash" "$squash".tmp.*
+            enroot import -o "${squash}.tmp.$$" "docker://$image"
+            mv -f "${squash}.tmp.$$" "$squash"
+        fi
+    ) || exit 1
+}
+
+import_squash "$SQUASH_FILE" "$IMAGE"
+import_squash "$NGINX_SQUASH_FILE" "$NGINX_IMAGE"
 
 export EVAL_ONLY="${EVAL_ONLY:-false}"
 
@@ -198,11 +227,12 @@ fi
 
 echo "Cloning srt-slurm repository..."
 SRT_REPO_DIR="srt-slurm"
+SRTCTL_SETUP_SCRIPT=""
 # On the watchtower (Oracle) gb200 cluster, /home/slurm-shared is not
 # cross-mounted to compute nodes. Put the srt-slurm workspace and staged
 # InferenceX checkout on a writable shared-FS path that compute can see.
 # Per-run-unique paths avoid races between parallel sweep jobs.
-if [[ $MODEL_PREFIX == "minimaxm2.5" || $MODEL_PREFIX == "kimik2.5" ]]; then
+if uses_watchtower_shared_fs; then
     SHARED_BASE=""
     for cand in \
         /mnt/lustre01/users-public/sa-shared/gha-runs \
@@ -274,6 +304,16 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm2.5" ]]; then
         echo "Unsupported minimaxm2.5 precision for GB200 dynamo-vllm: $PRECISION" >&2
         exit 1
     fi
+elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
+    cd "$SRT_REPO_DIR" || exit 1
+    git checkout sa-submission-q2-2026 || exit 1
+    mkdir -p recipes/vllm/minimax-m3-gb200-fp8 || exit 1
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8" recipes/vllm/minimax-m3-gb200-fp8 || exit 1
+    SRTCTL_SETUP_SCRIPT="minimax-m3-gb200-vllm-fixes.sh"
+    cp \
+        "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
+        "configs/$SRTCTL_SETUP_SCRIPT" || exit 1
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
     cd "$SRT_REPO_DIR" || exit 1
@@ -303,7 +343,7 @@ source $HOME/.local/bin/env
 # under a head-node-only path, .venv/bin/python3 becomes a broken
 # symlink on compute. Pin the venv to /usr/bin/python3 — a system
 # path that exists at the same location on both head and compute.
-if [[ ($MODEL_PREFIX == "minimaxm2.5" || $MODEL_PREFIX == "kimik2.5") && -x /usr/bin/python3 ]]; then
+if uses_watchtower_shared_fs && [[ -x /usr/bin/python3 ]]; then
     uv venv --seed --python /usr/bin/python3
 else
     uv venv --seed
@@ -320,10 +360,10 @@ echo "Configs available at: $SRT_REPO_DIR/"
 
 # Create srtslurm.yaml for srtctl (used by both frameworks)
 SRTCTL_ROOT="${GITHUB_WORKSPACE}/srt-slurm"
-# Minimax on watchtower: SRT_REPO_DIR was moved to a shared-FS path
+# Watchtower-hosted sweeps: SRT_REPO_DIR was moved to a shared-FS path
 # above so srtctl's outputs/ directory (which lives under
 # SRTCTL_ROOT) is visible to compute nodes.
-if [[ $MODEL_PREFIX == "minimaxm2.5" || $MODEL_PREFIX == "kimik2.5" ]]; then
+if uses_watchtower_shared_fs; then
     SRTCTL_ROOT="$SRT_REPO_DIR"
 fi
 echo "Creating srtslurm.yaml configuration..."
@@ -365,7 +405,7 @@ export INFMAX_WORKSPACE="$GITHUB_WORKSPACE"
 # can't see. Stage the relevant subset to shared FS and repoint
 # INFMAX_WORKSPACE there. rsync excludes the srt-slurm clone (already
 # on shared FS) and .git (not needed in container) for speed.
-if [[ $MODEL_PREFIX == "minimaxm2.5" || $MODEL_PREFIX == "kimik2.5" ]]; then
+if uses_watchtower_shared_fs; then
     SHARED_INFMAX_WORKSPACE="${SHARED_BASE}/infmax-workspace-${RUN_KEY}"
     mkdir -p "$SHARED_INFMAX_WORKSPACE" || exit 1
     rsync -a --delete \
@@ -390,11 +430,16 @@ if [[ ! -f "$CONFIG_PATH" ]]; then
 fi
 sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_PATH"
 
+SRTCTL_APPLY_ARGS=(
+    -f "$CONFIG_PATH"
+    --tags "gb200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
+)
 if [[ "$FRAMEWORK" == "dynamo-sglang" ]]; then
-    SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_PATH" --tags "gb200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" --setup-script install-torchao.sh 2>&1)
-else
-    SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_PATH" --tags "gb200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
+    SRTCTL_APPLY_ARGS+=(--setup-script install-torchao.sh)
+elif [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
+    SRTCTL_APPLY_ARGS+=(--setup-script "$SRTCTL_SETUP_SCRIPT")
 fi
+SRTCTL_OUTPUT=$(srtctl apply "${SRTCTL_APPLY_ARGS[@]}" 2>&1)
 echo "$SRTCTL_OUTPUT"
 
 JOB_ID=$(echo "$SRTCTL_OUTPUT" | grep -oP '✅ Job \K[0-9]+' || echo "$SRTCTL_OUTPUT" | grep -oP 'Job \K[0-9]+')