From d6e0efee0fd19f4a859592b28c52b55188c4a6d1 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 12:26:59 -0700
Subject: [PATCH 1/7] chore: route Kimi K2.6 sweep to b200-new
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 Kimi K2.6 B200 Dynamo vLLM 扫描切换到 b200-new，并移除 DEP4 到 DEP8、并发 1024 的配置项。
---
 .../disagg-b200-1p1d-dep4-dep8-c1024.yaml     | 112 ------------------
 configs/nvidia-master.yaml                    |  15 +--
 configs/runners.yaml                          |  10 ++
 perf-changelog.yaml                           |   8 ++
 runners/launch_b200-nscale-slurm.sh           |   3 +
 5 files changed, 22 insertions(+), 126 deletions(-)
 delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep4-dep8-c1024.yaml
 create mode 100644 runners/launch_b200-nscale-slurm.sh

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep4-dep8-c1024.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep4-dep8-c1024.yaml
deleted file mode 100644
index 5cb6c798b6..0000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep4-dep8-c1024.yaml
+++ /dev/null
@@ -1,112 +0,0 @@
-name: kimi-k2.6-vllm-disagg-b200-1p1d-dep4-dep8-c1024
-model:
-  path: kimi-k2.6-nvfp4
-  container: vllm/vllm-openai:v0.25.1
-  precision: fp4
-dynamo:
-  wheel: 1.3.0.dev20260721
-  install: true
-resources:
-  gpu_type: b200
-  gpus_per_node: 8
-  prefill_nodes: 1
-  decode_nodes: 1
-  prefill_workers: 1
-  decode_workers: 1
-  gpus_per_prefill: 4
-  gpus_per_decode: 8
-infra:
-  etcd_nats_dedicated_node: true
-frontend:
-  type: dynamo
-  enable_multiple_frontends: false
-backend:
-  type: vllm
-  connector: null
-  dp_launch_mode: per_gpu
-  prefill_environment:
-    PYTHONUNBUFFERED: '1'
-    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: '0'
-    VLLM_USE_FLASHINFER_MOE_FP4: '1'
-    VLLM_USE_NCCL_SYMM_MEM: '0'
-    NCCL_CUMEM_ENABLE: '1'
-    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: '900'
-    NCCL_WATCHDOG_TIMEOUT: '1800'
-    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
-    TORCH_CUDA_ARCH_LIST: '10.0'
-    PYTHONNOUSERSITE: '1'
-    UCX_MEMTYPE_CACHE: n
-    UCX_MEMTYPE_REG_WHOLE: n
-    UCX_CUDA_IPC_ENABLE_MNNVL: n
-    UCX_MAX_RMA_RAILS: '1'
-    UCX_MAX_RNDV_RAILS: '1'
-    UCX_RNDV_SCHEME: put_zcopy
-  decode_environment:
-    PYTHONUNBUFFERED: '1'
-    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: '0'
-    VLLM_USE_FLASHINFER_MOE_FP4: '1'
-    VLLM_USE_NCCL_SYMM_MEM: '0'
-    NCCL_CUMEM_ENABLE: '1'
-    NCCL_WATCHDOG_TIMEOUT: '1800'
-    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '1800'
-    TORCH_CUDA_ARCH_LIST: '10.0'
-    PYTHONNOUSERSITE: '1'
-    UCX_MEMTYPE_CACHE: n
-    UCX_MEMTYPE_REG_WHOLE: n
-    UCX_CUDA_IPC_ENABLE_MNNVL: n
-    UCX_MAX_RMA_RAILS: '1'
-    UCX_MAX_RNDV_RAILS: '1'
-    UCX_RNDV_SCHEME: put_zcopy
-  vllm_config:
-    prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
-      served-model-name: nvidia/Kimi-K2.6-NVFP4
-      kv-cache-dtype: fp8
-      tensor-parallel-size: 1
-      pipeline-parallel-size: 1
-      data-parallel-size: 4
-      data-parallel-rpc-port: 13346
-      enable-expert-parallel: true
-      max-model-len: 10240
-      max-num-seqs: 2048
-      enforce-eager: true
-      compilation-config: '{"custom_ops":["+quant_fp8","+rms_norm","+rotary_embedding"],"pass_config":{"fuse_attn_quant":true,"fuse_allreduce_rms":true}}'
-      max-num-batched-tokens: 8192
-      safetensors-load-strategy: prefetch
-      trust-remote-code: true
-      no-enable-prefix-caching: true
-      no-enable-flashinfer-autotune: true
-      attention-backend: FLASHINFER_MLA
-      block-size: 128
-      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED"}'
-      gpu-memory-utilization: 0.94
-    decode:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
-      served-model-name: nvidia/Kimi-K2.6-NVFP4
-      kv-cache-dtype: fp8
-      tensor-parallel-size: 1
-      pipeline-parallel-size: 1
-      data-parallel-size: 8
-      data-parallel-rpc-port: 13345
-      enable-expert-parallel: true
-      max-model-len: 10240
-      max-num-seqs: 2048
-      max-num-batched-tokens: 8192
-      safetensors-load-strategy: prefetch
-      trust-remote-code: true
-      no-enable-prefix-caching: true
-      no-enable-flashinfer-autotune: true
-      async-scheduling: true
-      attention-backend: FLASHINFER_MLA
-      block-size: 128
-      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","custom_ops":["+quant_fp8","+rms_norm","+rotary_embedding"],"pass_config":{"fuse_attn_quant":true,"fuse_allreduce_rms":true}}'
-      gpu-memory-utilization: 0.9
-      stream-interval: 50
-      max-cudagraph-capture-size: 1024
-benchmark:
-  type: sa-bench
-  isl: 8192
-  osl: 1024
-  concurrencies: '1024'
-  req_rate: inf
-  use_chat_template: true
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a30af59641..58c94561a9 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -4981,7 +4981,7 @@ kimik2.6-fp4-b200-dynamo-vllm:
   image: vllm/vllm-openai:v0.25.1
   model: nvidia/Kimi-K2.6-NVFP4
   model-prefix: kimik2.6
-  runner: b200-multinode
+  runner: b200-new
   precision: fp4
   framework: dynamo-vllm
   router: { name: dynamo-router, version: "1.3.0.dev20260721" }
@@ -5045,19 +5045,6 @@ kimik2.6-fp4-b200-dynamo-vllm:
           tp: 4
           ep: 1
           dp-attn: false
-      - conc-list: [1024]
-        prefill:
-          num-worker: 1
-          tp: 1
-          ep: 4
-          dp-attn: true
-          additional-settings:
-          - "CONFIG_FILE=recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep4-dep8-c1024.yaml"
-        decode:
-          num-worker: 1
-          tp: 1
-          ep: 8
-          dp-attn: true
       - conc-list: [2048]
         prefill:
           num-worker: 1
diff --git a/configs/runners.yaml b/configs/runners.yaml
index 5e992eb4e5..e6014c3dbe 100644
--- a/configs/runners.yaml
+++ b/configs/runners.yaml
@@ -80,6 +80,16 @@ labels:
     - b200-dgxc-slurm_7
     - b200-dgxc-slurm_8
     - b200-dgxc-slurm_9
+  b200-new:
+    - b200-nscale-slurm_0
+    - b200-nscale-slurm_1
+    - b200-nscale-slurm_2
+    - b200-nscale-slurm_3
+    - b200-nscale-slurm_4
+    - b200-nscale-slurm_5
+    - b200-nscale-slurm_6
+    - b200-nscale-slurm_7
+    - b200-nscale-slurm_8
   mi300x:
     - mi300x-amds_00
     - mi300x-amds_01
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6eebd97070..bd28eab357 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5355,3 +5355,11 @@
     - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path"
     - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213
+
+- config-keys:
+    - kimik2.6-fp4-b200-dynamo-vllm
+  description:
+    - "Run the Kimi K2.6 NVFP4 B200 Dynamo vLLM disaggregated configuration on the b200-new runner."
+    - "Remove the 1P/1D DEP4-to-DEP8 concurrency-1024 configuration."
+    - "Image: vllm/vllm-openai:v0.25.1"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
new file mode 100644
index 0000000000..184d1565fc
--- /dev/null
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -0,0 +1,3 @@
+#!/usr/bin/bash
+
+exec bash "$(dirname "${BASH_SOURCE[0]}")/launch_b200-dgxc.sh" "$@"

From 6f0341ec9d2484cc4de266b9585bdb15576f22f4 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 12:27:26 -0700
Subject: [PATCH 2/7] chore: update perf changelog PR link
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：更新性能变更日志中的拉取请求链接。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bd28eab357..f68e4635d3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5362,4 +5362,4 @@
     - "Run the Kimi K2.6 NVFP4 B200 Dynamo vLLM disaggregated configuration on the b200-new runner."
     - "Remove the 1P/1D DEP4-to-DEP8 concurrency-1024 configuration."
     - "Image: vllm/vllm-openai:v0.25.1"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2438

From 2550525c842ed17df143dc524c08e2b073ce8f38 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 12:53:00 -0700
Subject: [PATCH 3/7] fix: configure b200-new launch environment
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Configure the nscale Slurm account and partition, shared squash cache, model path, and runner temp storage while retaining the common B200 launcher defaults.

中文：配置 nscale 的 Slurm 账户和分区、共享 squash 缓存、模型路径以及 runner 临时存储，同时保留通用 B200 启动器的默认行为。
---
 runners/launch_b200-dgxc.sh         |  6 +++---
 runners/launch_b200-nscale-slurm.sh | 14 ++++++++++++++
 2 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index be68ce073b..c42d8b6f9f 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -1,8 +1,8 @@
 #!/usr/bin/bash
 
 # System-specific configuration for B200 DGXC Slurm cluster
-SLURM_PARTITION="gpu-2"
-SLURM_ACCOUNT="benchmark"
+SLURM_PARTITION="${SLURM_PARTITION:-gpu-2}"
+SLURM_ACCOUNT="${SLURM_ACCOUNT:-benchmark}"
 
 set -x
 
@@ -53,7 +53,7 @@ elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH="/lustre/fsw/models/Kimi-K2.5-NVFP4"
     export SRT_SLURM_MODEL_PREFIX="kimik2.5-fp4"
 elif [[ $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/lustre/fsw/models/Kimi-K2.6-NVFP4"
+    export MODEL_PATH="${MODEL_PATH:-/lustre/fsw/models/Kimi-K2.6-NVFP4}"
     export SRT_SLURM_MODEL_PREFIX="kimi-k2.6-nvfp4"
 elif [[ $MODEL_PREFIX == "minimaxm2.5" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH="/lustre/fsw/models/MiniMax-M2.5"
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 184d1565fc..6a998baea4 100644
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -1,3 +1,17 @@
 #!/usr/bin/bash
 
+export SLURM_PARTITION="batch_all"
+export SLURM_ACCOUNT="benchmark"
+export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"
+
+# The nscale login nodes can have a full node-local /tmp. Keep enroot's GNU
+# parallel buffers on runner-owned storage so image imports do not depend on
+# node-local temp capacity.
+export TMPDIR="${RUNNER_TEMP:-${GITHUB_WORKSPACE:-$PWD}/.runner-tmp}/inferencex-enroot"
+export B200_SQUASH_DIR="/data/home/sa-shared/containers"
+mkdir -p "$TMPDIR" "$B200_SQUASH_DIR" || {
+    echo "Unable to prepare nscale runner storage" >&2
+    exit 1
+}
+
 exec bash "$(dirname "${BASH_SOURCE[0]}")/launch_b200-dgxc.sh" "$@"

From 7eb3acb0f31de40bfaf557fc6aac818a351dd56b Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 13:12:30 -0700
Subject: [PATCH 4/7] fix: stabilize b200-new image staging
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use the shared squash cache with enroot local temporary storage and allow the first serialized image import enough time to finish.

中文：使用共享 squash 缓存和 enroot 本地临时存储，并为首次串行镜像导入预留足够的完成时间。
---
 runners/launch_b200-dgxc.sh         | 2 +-
 runners/launch_b200-nscale-slurm.sh | 7 ++-----
 2 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index c42d8b6f9f..28a0667f68 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -213,7 +213,7 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         local lock_file="${lock_dir}/${image_key}.lock"
 
         (
-            flock -w 600 9 || { echo "Failed to acquire lock for $squash_file" >&2; exit 1; }
+            flock -w "${B200_SQUASH_LOCK_TIMEOUT:-600}" 9 || { echo "Failed to acquire lock for $squash_file" >&2; exit 1; }
             if unsquashfs -l "$squash_file" > /dev/null 2>&1; then
                 echo "Squash file already exists and is valid, skipping import: $squash_file"
             else
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 6a998baea4..ce43cec4a8 100644
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -4,12 +4,9 @@ export SLURM_PARTITION="batch_all"
 export SLURM_ACCOUNT="benchmark"
 export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"
 
-# The nscale login nodes can have a full node-local /tmp. Keep enroot's GNU
-# parallel buffers on runner-owned storage so image imports do not depend on
-# node-local temp capacity.
-export TMPDIR="${RUNNER_TEMP:-${GITHUB_WORKSPACE:-$PWD}/.runner-tmp}/inferencex-enroot"
 export B200_SQUASH_DIR="/data/home/sa-shared/containers"
-mkdir -p "$TMPDIR" "$B200_SQUASH_DIR" || {
+export B200_SQUASH_LOCK_TIMEOUT="3600"
+mkdir -p "$B200_SQUASH_DIR" || {
     echo "Unable to prepare nscale runner storage" >&2
     exit 1
 }

From 938b3806a8491776774118437793582c784358bd Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 14:59:45 -0700
Subject: [PATCH 5/7] fix: use batch_1 for b200-new
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 b200-new 启动器的 Slurm 分区从 batch_all 更新为 batch_1。
---
 runners/launch_b200-nscale-slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index ce43cec4a8..de54528e11 100644
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/bash
 
-export SLURM_PARTITION="batch_all"
+export SLURM_PARTITION="batch_1"
 export SLURM_ACCOUNT="benchmark"
 export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"
 

From 57faf6db0d97084364e7788e42bdd95a63df1fa5 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 15:17:07 -0700
Subject: [PATCH 6/7] fix: use batch_2 for b200-new
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 b200-new 启动器的 Slurm 分区更新为 batch_2。
---
 runners/launch_b200-nscale-slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index de54528e11..955ed36354 100644
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/bash
 
-export SLURM_PARTITION="batch_1"
+export SLURM_PARTITION="batch_2"
 export SLURM_ACCOUNT="benchmark"
 export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"
 

From fc2609111631afc2363f1133fad5a960e9638845 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 31 Jul 2026 15:19:34 -0700
Subject: [PATCH 7/7] revert: restore batch_1 for b200-new
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Restore the b200-new Slurm partition from batch_2 to batch_1.\n\n中文：将 b200-new 的 Slurm 分区从 batch_2 恢复为 batch_1。
---
 runners/launch_b200-nscale-slurm.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 955ed36354..de54528e11 100644
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/bash
 
-export SLURM_PARTITION="batch_2"
+export SLURM_PARTITION="batch_1"
 export SLURM_ACCOUNT="benchmark"
 export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"