diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 96a68b589..41e694489 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -12035,10 +12035,9 @@ minimaxm3-fp8-b300-dynamo-vllm:
           ep: 8
           dp-attn: false
 
-# MiniMax-M3 GB300 disagg sweep — adapted from NV B300 PR #1863.
-# All prefill DEP2 (TP1 DP2 EP, 2 GPU/worker). Decode: TP4+Marlin, TEP8,
-# DEP8, DEP4. 4 GPU/node (GB300 NVL72). 4p3d (3 decode workers) skipped.
-# kv-cache-dtype=fp8 added. srun_options mem=0 required.
+# MiniMax-M3 GB300 disagg sweep — refreshed recipe set (no Marlin variants).
+# All prefill DEP2 (TP1 DP2 EP, 2 GPU/worker). Decode: DEP4, TEP8, DEP8, TEP4.
+# 4 GPU/node (GB300 NVL72). kv-cache-dtype=fp8. srun_options mem=0 required.
 minimaxm3-fp8-gb300-dynamo-vllm:
   image: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223
   model: MiniMaxAI/MiniMax-M3-MXFP8
@@ -12053,155 +12052,155 @@ minimaxm3-fp8-gb300-dynamo-vllm:
     - isl: 1024
       osl: 1024
       search-space:
-      # 1p1d DEP2+TEP8, 3n: conc 4,16,64,128,4096
-      - conc-list: [4, 16, 64, 128, 4096]
+      # 1p1d DEP2+DEP4, 2n: conc 8192
+      - conc-list: [8192]
         prefill:
           num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tep8-3n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml"
         decode:
           num-worker: 1
-          tp: 8
-          ep: 8
-          dp-attn: false
+          tp: 4
+          ep: 4
+          dp-attn: true
 
-      # 1p1d DEP2+TP4 Marlin, 2n: conc 1,4,8,16
-      - conc-list: [1, 4, 8, 16]
+      # 1p2d DEP2+TEP8, 5n: conc 4,16,64,128,256
+      - conc-list: [4, 16, 64, 128, 256]
         prefill:
           num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml"
         decode:
-          num-worker: 1
-          tp: 4
-          ep: 1
+          num-worker: 2
+          tp: 8
+          ep: 8
           dp-attn: false
 
-      # 1p2d DEP2+DEP4, 3n: conc 2048
-      - conc-list: [2048]
+      # 2p2d DEP2+TEP8, 5n: conc 32
+      - conc-list: [32]
         prefill:
-          num-worker: 1
+          num-worker: 2
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p2d-dep2-dep4-3n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml"
         decode:
           num-worker: 2
-          tp: 4
-          ep: 4
-          dp-attn: true
+          tp: 8
+          ep: 8
+          dp-attn: false
 
-      # 2p1d DEP2+DEP8, 3n: conc 512,4096
-      - conc-list: [512, 4096]
+      # 2p3d DEP2+DEP4, 4n: conc 8192
+      - conc-list: [8192]
         prefill:
           num-worker: 2
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-dep8-3n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml"
         decode:
-          num-worker: 1
-          tp: 8
-          ep: 8
+          num-worker: 3
+          tp: 4
+          ep: 4
           dp-attn: true
 
-      # 2p1d DEP2+TEP8, 3n: conc 32
-      - conc-list: [32]
+      # 2p4d DEP2+DEP4, 5n: conc 8192
+      - conc-list: [8192]
         prefill:
           num-worker: 2
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-tep8-3n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml"
         decode:
-          num-worker: 1
-          tp: 8
-          ep: 8
-          dp-attn: false
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: true
 
-      # 2p2d DEP2+TEP8, 5n: conc 16
-      - conc-list: [16]
+      # 4p2d DEP2+DEP8, 6n: conc 1024,4096
+      - conc-list: [1024, 4096]
         prefill:
-          num-worker: 2
+          num-worker: 4
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml"
         decode:
           num-worker: 2
           tp: 8
           ep: 8
-          dp-attn: false
+          dp-attn: true
 
-      # 3p2d DEP2+TEP8, 6n: conc 4
-      - conc-list: [4]
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1p1d DEP2+DEP8, 3n: conc 256
+      - conc-list: [256]
         prefill:
-          num-worker: 3
+          num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml"
         decode:
-          num-worker: 2
+          num-worker: 1
           tp: 8
           ep: 8
-          dp-attn: false
+          dp-attn: true
 
-    - isl: 8192
-      osl: 1024
-      search-space:
-      # 1p1d DEP2+TP4 Marlin, 2n: conc 1,4,8,16
-      - conc-list: [1, 4, 8, 16]
+      # 1p1d DEP2+TEP8, 3n: conc 128
+      - conc-list: [128]
         prefill:
           num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml"
         decode:
           num-worker: 1
-          tp: 4
-          ep: 1
+          tp: 8
+          ep: 8
           dp-attn: false
 
-      # 1p2d DEP2+DEP8, 5n: conc 128
-      - conc-list: [128]
+      # 1p2d DEP2+TEP8, 5n: conc 32,64,128
+      - conc-list: [32, 64, 128]
         prefill:
           num-worker: 1
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p2d-dep2-dep8-5n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml"
         decode:
           num-worker: 2
           tp: 8
           ep: 8
-          dp-attn: true
+          dp-attn: false
 
-      # 2p2d DEP2+DEP8, 5n: conc 256,512
-      - conc-list: [256, 512]
+      # 2p1d DEP2+DEP8, 3n: conc 512
+      - conc-list: [512]
         prefill:
           num-worker: 2
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-dep8-5n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml"
         decode:
-          num-worker: 2
+          num-worker: 1
           tp: 8
           ep: 8
           dp-attn: true
@@ -12214,72 +12213,72 @@ minimaxm3-fp8-gb300-dynamo-vllm:
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml"
         decode:
           num-worker: 2
           tp: 8
           ep: 8
           dp-attn: false
 
-      # 3p2d DEP2+DEP8, 6n: conc 512
-      - conc-list: [512]
+      # 2p4d DEP2+TEP4, 5n: conc 4
+      - conc-list: [4]
         prefill:
-          num-worker: 3
+          num-worker: 2
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-dep8-6n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml"
         decode:
-          num-worker: 2
-          tp: 8
-          ep: 8
-          dp-attn: true
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: false
 
-      # 3p2d DEP2+TEP8, 6n: conc 32
-      - conc-list: [32]
+      # 3p1d DEP2+DEP8, 4n: conc 1024
+      - conc-list: [1024]
         prefill:
           num-worker: 3
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml"
         decode:
-          num-worker: 2
+          num-worker: 1
           tp: 8
           ep: 8
-          dp-attn: false
+          dp-attn: true
 
-      # 4p2d DEP2+DEP8, 6n: conc 4096
-      - conc-list: [4096]
+      # 3p2d DEP2+DEP8, 6n: conc 512
+      - conc-list: [512]
         prefill:
-          num-worker: 4
+          num-worker: 3
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-4p2d-dep2-dep8-6n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml"
         decode:
           num-worker: 2
           tp: 8
           ep: 8
           dp-attn: true
 
-      # 5p2d DEP2+TEP8, 7n: conc 4,64
-      - conc-list: [4, 64]
+      # 6p1d DEP2+DEP8, 5n: conc 2048
+      - conc-list: [2048]
         prefill:
-          num-worker: 5
+          num-worker: 6
           tp: 2
           ep: 2
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-5p2d-dep2-tep8-7n.yaml"
+          - "CONFIG_FILE=recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml"
         decode:
-          num-worker: 2
+          num-worker: 1
           tp: 8
           ep: 8
-          dp-attn: false
+          dp-attn: true
 
 qwen3.5-fp4-b200-trt:
   image: nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc18
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p2d-dep2-dep4-3n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
similarity index 76%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p2d-dep2-dep4-3n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
index af5315c76..f57d7af09 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p2d-dep2-dep4-3n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-dep4-fp8-1k1k"
+name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-dep4-fp8-1k1k"
 
-# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP4 (TP1 DP4 EP)
-# decode = 3 nodes (1P + 2D). Adapted from NV B300 PR #1863.
+# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP4 decode (TP1 DP4 EP, 4 GPU/worker = 1 node each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# Nodes: 1 prefill + 1 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 1
-  decode_nodes: 2
+  decode_nodes: 1
   prefill_workers: 1
-  decode_workers: 2
+  decode_workers: 1
   gpus_per_prefill: 2
   gpus_per_decode: 4
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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: 4
       data-parallel-rpc-port: 13345
@@ -92,8 +94,8 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -101,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 1024
   osl: 1024
-  concurrencies: "2048"
+  concurrencies: "8192"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tep8-3n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
similarity index 77%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tep8-3n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
index 4b00b5660..b4f457654 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tep8-3n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tep8-fp8-1k1k"
+name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-tep8-fp8-1k1k"
 
-# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D TEP8 (TP8+EP8)
-# decode = 3 nodes (1P + 2D). Adapted from NV B300 PR #1863.
+# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 1
-  decode_nodes: 2
+  decode_nodes: 4
   prefill_workers: 1
-  decode_workers: 1
+  decode_workers: 2
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -81,6 +82,7 @@ backend:
 
     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
@@ -93,11 +95,11 @@ backend:
       stream-interval: 32
       max-num-seqs: 4096
       max-num-batched-tokens: 16384
-      max-cudagraph-capture-size: 8192
+      max-cudagraph-capture-size: 8196
 
 benchmark:
   type: "sa-bench"
   isl: 1024
   osl: 1024
-  concurrencies: "4x16x64x128x4096"
+  concurrencies: "4x16x64x128x256"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
similarity index 81%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
index 5babf0835..6bba9ea86 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
@@ -1,30 +1,30 @@
 name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-tep8-fp8-1k1k"
 
-# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 (TP8+EP8)
-# decode = 5 nodes (1P + 4D). Adapted from NV B300 PR #1863.
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -81,6 +82,7 @@ backend:
 
     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
@@ -93,11 +95,11 @@ backend:
       stream-interval: 32
       max-num-seqs: 4096
       max-num-batched-tokens: 16384
-      max-cudagraph-capture-size: 8192
+      max-cudagraph-capture-size: 8196
 
 benchmark:
   type: "sa-bench"
   isl: 1024
   osl: 1024
-  concurrencies: "16"
+  concurrencies: "32"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
similarity index 71%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
index 26fa89b94..de852e427 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tp4-marlin-fp8-1k1k"
+name: "minimax-m3-vllm-disagg-gb300-2p3d-dep2-dep4-fp8-1k1k"
 
-# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D TP4 Marlin
-# decode = 2 nodes (1P + 1D). Adapted from NV B300 PR #1863.
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 3D DEP4 decode (TP1 DP4 EP, 4 GPU/worker = 1 node each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# Nodes: 1 prefill + 3 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 1
-  decode_nodes: 1
-  prefill_workers: 1
-  decode_workers: 1
+  decode_nodes: 3
+  prefill_workers: 2
+  decode_workers: 3
   gpus_per_prefill: 2
   gpus_per_decode: 4
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,15 +76,17 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      stream-interval: 100
       max-cudagraph-capture-size: 2048
       max-num-batched-tokens: 2048
 
     decode:
       kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
-      tensor-parallel-size: 4
-      enable-expert-parallel: false
-      moe-backend: marlin
+      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
@@ -91,8 +94,8 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 4096
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -100,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 1024
   osl: 1024
-  concurrencies: "1x4x8x16"
+  concurrencies: "8192"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml
new file mode 100644
index 000000000..8f7b7b140
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb300-2p4d-dep2-dep4-fp8-1k1k"
+
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 4D DEP4 decode (TP1 DP4 EP, 4 GPU/worker = 1 node each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 2
+  decode_workers: 4
+  gpus_per_prefill: 2
+  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: 2
+      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: 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-num-seqs: 1024
+      max-num-batched-tokens: 16384
+      max-cudagraph-capture-size: 2048
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "8192"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-dep8-3n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
similarity index 75%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-dep8-3n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
index 7cc5f50c4..f6cf6a59f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-dep8-3n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-2p1d-dep2-dep8-fp8-1k1k"
+name: "minimax-m3-vllm-disagg-gb300-4p2d-dep2-dep8-fp8-1k1k"
 
-# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP8 (TP1 DP8 EP)
-# decode = 3 nodes (1P + 2D). Adapted from NV B300 PR #1863.
+# 4P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
-  prefill_nodes: 1
-  decode_nodes: 2
-  prefill_workers: 2
-  decode_workers: 1
+  prefill_nodes: 2
+  decode_nodes: 4
+  prefill_workers: 4
+  decode_workers: 2
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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
@@ -92,8 +94,8 @@ backend:
       max-model-len: 2304
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 4096
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -101,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 1024
   osl: 1024
-  concurrencies: "512x4096"
+  concurrencies: "1024x4096"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-tep8-3n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-tep8-3n.yaml
deleted file mode 100644
index 0c4f3498c..000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-2p1d-dep2-tep8-3n.yaml
+++ /dev/null
@@ -1,103 +0,0 @@
-name: "minimax-m3-vllm-disagg-gb300-2p1d-dep2-tep8-fp8-1k1k"
-
-# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D TEP8 (TP8+EP8)
-# decode = 3 nodes (1P + 2D). Adapted from NV B300 PR #1863.
-
-model:
-  path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
-  precision: "fp8"
-
-dynamo:
-  install: true
-  wheel: "1.2.0.dev20260526"
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "72"
-srun_options:
-  mem: "0"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
-resources:
-  gpu_type: "gb300"
-  gpus_per_node: 4
-  prefill_nodes: 1
-  decode_nodes: 2
-  prefill_workers: 2
-  decode_workers: 1
-  gpus_per_prefill: 2
-  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"}'
-      tensor-parallel-size: 1
-      data-parallel-size: 2
-      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"}'
-      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: 8192
-
-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-gb300-fp8/1k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml
deleted file mode 100644
index d4176055a..000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml
+++ /dev/null
@@ -1,103 +0,0 @@
-name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-tep8-fp8-1k1k"
-
-# 3P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 (TP8+EP8)
-# decode = 6 nodes (2P + 4D). Adapted from NV B300 PR #1863.
-
-model:
-  path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
-  precision: "fp8"
-
-dynamo:
-  install: true
-  wheel: "1.2.0.dev20260526"
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "72"
-srun_options:
-  mem: "0"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
-resources:
-  gpu_type: "gb300"
-  gpus_per_node: 4
-  prefill_nodes: 2
-  decode_nodes: 4
-  prefill_workers: 3
-  decode_workers: 2
-  gpus_per_prefill: 2
-  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"}'
-      tensor-parallel-size: 1
-      data-parallel-size: 2
-      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"}'
-      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: 8192
-
-benchmark:
-  type: "sa-bench"
-  isl: 1024
-  osl: 1024
-  concurrencies: "4"
-  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p2d-dep2-dep8-5n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
similarity index 76%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p2d-dep2-dep8-5n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
index b56b65b26..d990d661b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p2d-dep2-dep8-5n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-dep8-fp8-8k1k"
+name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-dep8-fp8-8k1k"
 
-# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 (TP1 DP8 EP)
-# decode = 5 nodes (1P + 4D). Adapted from NV B300 PR #1863.
+# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 1
-  decode_nodes: 4
+  decode_nodes: 2
   prefill_workers: 1
-  decode_workers: 2
+  decode_workers: 1
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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
@@ -92,8 +94,8 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -101,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 8192
   osl: 1024
-  concurrencies: "128"
+  concurrencies: "256"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
similarity index 77%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
index 35950dc32..d46133924 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-tep8-6n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-tep8-fp8-8k1k"
+name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tep8-fp8-8k1k"
 
-# 3P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 (TP8+EP8)
-# decode = 6 nodes (2P + 4D). Adapted from NV B300 PR #1863.
+# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
-  prefill_nodes: 2
-  decode_nodes: 4
-  prefill_workers: 3
-  decode_workers: 2
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -81,6 +82,7 @@ backend:
 
     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
@@ -99,5 +101,5 @@ benchmark:
   type: "sa-bench"
   isl: 8192
   osl: 1024
-  concurrencies: "32"
+  concurrencies: "128"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-5p2d-dep2-tep8-7n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
similarity index 78%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-5p2d-dep2-tep8-7n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
index dbc9c5c9a..e8c606e27 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-5p2d-dep2-tep8-7n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
@@ -1,36 +1,36 @@
-name: "minimax-m3-vllm-disagg-gb300-5p2d-dep2-tep8-fp8-8k1k"
+name: "minimax-m3-vllm-disagg-gb300-1p2d-dep2-tep8-fp8-8k1k"
 
-# 5P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 (TP8+EP8)
-# decode = 7 nodes (3P + 4D). Adapted from NV B300 PR #1863.
+# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
-  prefill_nodes: 3
+  prefill_nodes: 1
   decode_nodes: 4
-  prefill_workers: 5
+  prefill_workers: 1
   decode_workers: 2
   gpus_per_prefill: 2
   gpus_per_decode: 8
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -81,6 +82,7 @@ backend:
 
     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
@@ -99,5 +101,5 @@ benchmark:
   type: "sa-bench"
   isl: 8192
   osl: 1024
-  concurrencies: "4x64"
+  concurrencies: "32x64x128"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-dep8-5n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
similarity index 76%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-dep8-5n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
index 7beba3420..02c3be14a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-dep8-5n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-dep8-fp8-8k1k"
+name: "minimax-m3-vllm-disagg-gb300-2p1d-dep2-dep8-fp8-8k1k"
 
-# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 (TP1 DP8 EP)
-# decode = 5 nodes (1P + 4D). Adapted from NV B300 PR #1863.
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 1
-  decode_nodes: 4
+  decode_nodes: 2
   prefill_workers: 2
-  decode_workers: 2
+  decode_workers: 1
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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
@@ -92,8 +94,8 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -101,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 8192
   osl: 1024
-  concurrencies: "256x512"
+  concurrencies: "512"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
similarity index 83%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
index 1ea678ace..304650d6c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-2p2d-dep2-tep8-5n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
@@ -1,30 +1,30 @@
 name: "minimax-m3-vllm-disagg-gb300-2p2d-dep2-tep8-fp8-8k1k"
 
-# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 (TP8+EP8)
-# decode = 5 nodes (1P + 4D). Adapted from NV B300 PR #1863.
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D TEP8 decode (TP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -81,6 +82,7 @@ backend:
 
     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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml
new file mode 100644
index 000000000..efea8bfac
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml
@@ -0,0 +1,105 @@
+name: "minimax-m3-vllm-disagg-gb300-2p4d-dep2-tep4-fp8-8k1k"
+
+# 2P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 4D TEP4 decode (TP4 EP, 4 GPU/worker = 1 node each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 2
+  decode_workers: 4
+  gpus_per_prefill: 2
+  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: 2
+      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-gb300-fp8/8k1k/disagg-gb300-4p2d-dep2-dep8-6n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
similarity index 75%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-4p2d-dep2-dep8-6n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
index 1526cd7ad..97e1ec88c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-4p2d-dep2-dep8-6n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
@@ -1,37 +1,37 @@
-name: "minimax-m3-vllm-disagg-gb300-4p2d-dep2-dep8-fp8-8k1k"
+name: "minimax-m3-vllm-disagg-gb300-3p1d-dep2-dep8-fp8-8k1k"
 
-# 4P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 (TP1 DP8 EP)
-# decode = 6 nodes (2P + 4D). Adapted from NV B300 PR #1863.
+# 3P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# Nodes: 2 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
   prefill_nodes: 2
-  decode_nodes: 4
-  prefill_workers: 4
-  decode_workers: 2
+  decode_nodes: 2
+  prefill_workers: 3
+  decode_workers: 1
   gpus_per_prefill: 2
   gpus_per_decode: 8
 
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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
@@ -92,8 +94,8 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
@@ -101,5 +103,5 @@ benchmark:
   type: "sa-bench"
   isl: 8192
   osl: 1024
-  concurrencies: "4096"
+  concurrencies: "1024"
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-dep8-6n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
similarity index 81%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-dep8-6n.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
index f4e000a5f..745b2fad4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-3p2d-dep2-dep8-6n.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
@@ -1,30 +1,30 @@
 name: "minimax-m3-vllm-disagg-gb300-3p2d-dep2-dep8-fp8-8k1k"
 
-# 3P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 (TP1 DP8 EP)
-# decode = 6 nodes (2P + 4D). Adapted from NV B300 PR #1863.
+# 3P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 2D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# 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
-  wheel: "1.2.0.dev20260526"
+  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"
 
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
 resources:
   gpu_type: "gb300"
   gpus_per_node: 4
@@ -64,6 +64,7 @@ backend:
   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: 2
       data-parallel-rpc-port: 13345
@@ -75,12 +76,13 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
+      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
@@ -92,8 +94,8 @@ backend:
       max-model-len: 9472
       language-model-only: true
       kv-cache-dtype: fp8
-      stream-interval: 32
-      max-num-seqs: 512
+      stream-interval: 100
+      max-num-seqs: 1024
       max-num-batched-tokens: 16384
       max-cudagraph-capture-size: 2048
 
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml
new file mode 100644
index 000000000..9be5cc177
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml
@@ -0,0 +1,107 @@
+name: "minimax-m3-vllm-disagg-gb300-6p1d-dep2-dep8-fp8-8k1k"
+
+# 6P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D DEP8 decode (TP1 DP8 EP, 8 GPU/worker = 2 nodes each).
+# GB300 has 4 GPUs/node. Adapted from NV B300 PR #1863.
+# Nodes: 3 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: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 3
+  decode_nodes: 2
+  prefill_workers: 6
+  decode_workers: 1
+  gpus_per_prefill: 2
+  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: 2
+      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: "2048"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml
deleted file mode 100644
index 4ee41241e..000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/disagg-gb300-1p1d-dep2-tp4-marlin-2n.yaml
+++ /dev/null
@@ -1,104 +0,0 @@
-name: "minimax-m3-vllm-disagg-gb300-1p1d-dep2-tp4-marlin-fp8-8k1k"
-
-# 1P DEP2 prefill (TP1 DP2 EP, 2 GPU/worker) + 1D TP4 Marlin
-# decode = 2 nodes (1P + 1D). Adapted from NV B300 PR #1863.
-
-model:
-  path: "minimax-m3-mxfp8"
-  container: "vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223"
-  precision: "fp8"
-
-dynamo:
-  install: true
-  wheel: "1.2.0.dev20260526"
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "72"
-srun_options:
-  mem: "0"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 720
-  interval_seconds: 10
-
-resources:
-  gpu_type: "gb300"
-  gpus_per_node: 4
-  prefill_nodes: 1
-  decode_nodes: 1
-  prefill_workers: 1
-  decode_workers: 1
-  gpus_per_prefill: 2
-  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"}'
-      tensor-parallel-size: 1
-      data-parallel-size: 2
-      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"}'
-      tensor-parallel-size: 4
-      enable-expert-parallel: false
-      moe-backend: marlin
-      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: 2048
-
-benchmark:
-  type: "sa-bench"
-  isl: 8192
-  osl: 1024
-  concurrencies: "1x4x8x16"
-  req_rate: "inf"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b3c7d283a..0ce7ceecd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4222,6 +4222,15 @@
     - "Initial submission: MiniMax-M3 MXFP4 disagg (prefill/decode) on MI355X with vLLM over the MoRI-IO KV connector (8k/1k)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1914
 
+- config-keys:
+    - minimaxm3-fp8-gb300-dynamo-vllm
+  description:
+    - "Refresh GB300 MiniMax-M3 disagg recipe set: replace disagg-gb300-* files with new naming convention; drop TP4+Marlin variants."
+    - "1k/1k topologies (6 shapes): 1p1d-dep2-dep4 (conc 8192), 1p2d-dep2-tep8 (conc 4,16,64,128,256), 2p2d-dep2-tep8 (conc 32), 2p3d-dep2-dep4 (conc 8192), 2p4d-dep2-dep4 (conc 8192), 4p2d-dep2-dep8 (conc 1024,4096)."
+    - "8k/1k topologies (9 shapes): 1p1d-dep2-dep8 (conc 256), 1p1d-dep2-tep8 (conc 128), 1p2d-dep2-tep8 (conc 32,64,128), 2p1d-dep2-dep8 (conc 512), 2p2d-dep2-tep8 (conc 16), 2p4d-dep2-tep4 (conc 4), 3p1d-dep2-dep8 (conc 1024), 3p2d-dep2-dep8 (conc 512), 6p1d-dep2-dep8 (conc 2048)."
+    - "Image unchanged: vllm/vllm-openai:minimax-m3-perf-arm64-13.0.1-7a67223."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1925
+
 - config-keys:
     - minimaxm3-fp4-b300-vllm
   description:
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 689451443..676615500 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -182,7 +182,7 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm2.5" && $PRECIS
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
-    git checkout main
+    git checkout sa-submission-q2-2026
     mkdir -p recipes/vllm/minimax-m3-gb300-fp8
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8" recipes/vllm/minimax-m3-gb300-fp8
     SRTCTL_SETUP_SCRIPT="minimax-m3-gb300-vllm-fixes.sh"