From 58a5304880eb4d16b7b8bda390500956a9fe9ded Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 13 Jul 2026 15:51:00 -0700
Subject: [PATCH 1/5] feat: add Kimi-K2.6 NVFP4 B300 disagg multinode
 dynamo-vllm benchmarks
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add kimik2.6-fp4-b300-dynamo-vllm to configs/nvidia-master.yaml with seven
8k1k disaggregated topologies, stage the vLLM recipes under
benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/, wire the
model path and a pinned srt-slurm checkout into runners/launch_b300-nv.sh
(with --no-preflight for the compute-node-local model path), and append the
perf-changelog entry.

中文：新增 Kimi-K2.6 NVFP4 B300 分离式多节点 dynamo-vllm 基准测试：在
configs/nvidia-master.yaml 中新增 kimik2.6-fp4-b300-dynamo-vllm（七个 8k1k
分离式拓扑），在 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/
下新增 vLLM 配方文件，在 runners/launch_b300-nv.sh 中接入模型路径与固定
commit 的 srt-slurm 检出（模型路径仅存在于计算节点本地，故使用
--no-preflight），并追加 perf-changelog 记录。
---
 .../disagg-b300-1p10d-dep4-tp4-atune.yaml     |  88 +++++++++++++++
 .../8k1k/disagg-b300-1p1d-dep4-dep8.yaml      |  91 +++++++++++++++
 .../8k1k/disagg-b300-1p8d-dep4-tp4.yaml       |  88 +++++++++++++++
 .../8k1k/disagg-b300-2p5d-dep4-tp8.yaml       |  89 +++++++++++++++
 .../8k1k/disagg-b300-3p1d-dep4-dep8.yaml      |  91 +++++++++++++++
 .../disagg-b300-4p1d-dep4-dep8-atune.yaml     |  91 +++++++++++++++
 .../8k1k/disagg-b300-4p1d-dep4-dep8.yaml      |  91 +++++++++++++++
 configs/nvidia-master.yaml                    | 105 ++++++++++++++++++
 perf-changelog.yaml                           |   8 ++
 runners/launch_b300-nv.sh                     |  21 +++-
 10 files changed, 762 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
new file mode 100644
index 0000000000..d1ce703229
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
@@ -0,0 +1,88 @@
+name: kimi-k2.6-vllm-disagg-b300-1p10d-dep4-tp4-atune
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 5
+  prefill_workers: 1
+  decode_workers: 10
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+infra:
+  etcd_nats_dedicated_node: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      safetensors-load-strategy: prefetch
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      enable-flashinfer-autotune: true
+      attention-backend: FLASHINFER_MLA
+      block-size: 128
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED"}'
+      gpu-memory-utilization: 0.9
+    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: 4
+      pipeline-parallel-size: 1
+      max-model-len: 9216
+      max-num-seqs: 2048
+      max-num-batched-tokens: 8192
+      safetensors-load-strategy: prefetch
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      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: 0.93
+      stream-interval: 50
+      max-cudagraph-capture-size: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
new file mode 100644
index 0000000000..7e2471d7a9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
@@ -0,0 +1,91 @@
+name: kimi-k2.6-vllm-disagg-b300-1p1d-dep4-dep8
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  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
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      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: 0.9
+    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: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
new file mode 100644
index 0000000000..965b118ffd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
@@ -0,0 +1,88 @@
+name: kimi-k2.6-vllm-disagg-b300-1p8d-dep4-tp4
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 8
+  gpus_per_prefill: 4
+  gpus_per_decode: 4
+infra:
+  etcd_nats_dedicated_node: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      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: 0.9
+    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: 4
+      pipeline-parallel-size: 1
+      max-model-len: 9216
+      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: 0.93
+      stream-interval: 50
+      max-cudagraph-capture-size: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
new file mode 100644
index 0000000000..7f5297121f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
@@ -0,0 +1,89 @@
+name: kimi-k2.6-vllm-disagg-b300-2p5d-dep4-tp8
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 5
+  prefill_workers: 2
+  decode_workers: 5
+  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
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "0"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      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: 0.9
+    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: 8
+      pipeline-parallel-size: 1
+      disable-custom-all-reduce: true
+      max-model-len: 9216
+      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: 0.93
+      stream-interval: 50
+      max-cudagraph-capture-size: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
new file mode 100644
index 0000000000..847bef8744
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
@@ -0,0 +1,91 @@
+name: kimi-k2.6-vllm-disagg-b300-3p1d-dep4-dep8
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 3
+  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
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      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: 0.9
+    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: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
new file mode 100644
index 0000000000..edce630d10
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
@@ -0,0 +1,91 @@
+name: kimi-k2.6-vllm-disagg-b300-4p1d-dep4-dep8-atune
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 4
+  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
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      safetensors-load-strategy: prefetch
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      enable-flashinfer-autotune: true
+      attention-backend: FLASHINFER_MLA
+      block-size: 128
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED"}'
+      gpu-memory-utilization: 0.9
+    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
+      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: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
new file mode 100644
index 0000000000..93401c50f9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
@@ -0,0 +1,91 @@
+name: kimi-k2.6-vllm-disagg-b300-4p1d-dep4-dep8
+model:
+  path: kimi-k2.6-nvfp4
+  container: vllm/vllm-openai:v0.22.0
+  precision: fp4
+dynamo:
+  version: 1.2.0.dev20260529
+  install: true
+setup_script: vllm-container-deps.sh
+resources:
+  gpu_type: b300
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 4
+  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
+  prefill_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "900"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  decode_environment:
+    VLLM_USE_FLASHINFER_MOE_FP4: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_WATCHDOG_TIMEOUT: "1800"
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
+  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: 4096
+      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: 32768
+      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: 0.9
+    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: 2048
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x4x16x64x128x256x512x1024x2048x3072x4096"
+  req_rate: "inf"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 56ac8f70f9..94405e00d8 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -14030,6 +14030,111 @@ kimik2.5-fp4-gb300-dynamo-vllm:
           tp: 1
           ep: 24
           dp-attn: true
+kimik2.6-fp4-b300-dynamo-vllm:
+  image: vllm/vllm-openai:v0.22.0
+  model: nvidia/Kimi-K2.6-NVFP4
+  model-prefix: kimik2.6
+  runner: b300
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 3
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 4
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - conc-list: [4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 4
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml"
+        decode:
+          num-worker: 8
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml"
+        decode:
+          num-worker: 10
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - conc-list: [1, 4, 16, 64, 128, 256, 512, 1024, 2048, 3072, 4096]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml"
+        decode:
+          num-worker: 5
+          tp: 8
+          ep: 1
+          dp-attn: false
 minimaxm3-fp8-h100-vllm-agentic:
   image: vllm/vllm-openai:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 57d510dd15..f850b543e2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4750,3 +4750,11 @@
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928"
     - "6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2137
+
+- config-keys:
+    - kimik2.6-fp4-b300-dynamo-vllm
+  description:
+    - "Add Kimi-K2.6 NVFP4 B300 disaggregated multinode vLLM benchmarks via Dynamo"
+    - "Image: vllm/vllm-openai:v0.22.0"
+    - "7 topologies at 8k/1k: 1p1d/3p1d/4p1d dep4-dep8 (4p1d also as FlashInfer-autotune variant at conc 4-4096), 1p8d dep4-tp4, 1p10d dep4-tp4 (FlashInfer autotune), 2p5d dep4-tp8; conc 1-4096; recipes under benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 6810ee5d85..2fc67e3d02 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -53,8 +53,13 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && $FRAMEWORK == "dy
 elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" && $FRAMEWORK == "dynamo-vllm" ]]; then
     export MODEL_PATH="/data/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
+elif [[ $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-vllm" ]]; then
+    # Node-local NVMe path; only exists on compute nodes, so srtctl apply
+    # runs with --no-preflight for this model (see SRTCTL_APPLY_ARGS below).
+    export MODEL_PATH="/scratch/models/Kimi-K2.6-NVFP4"
+    export SRT_SLURM_MODEL_PREFIX="kimi-k2.6-nvfp4"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm, kimik2.6-fp4 with dynamo-vllm"
     exit 1
 fi
 
@@ -92,6 +97,13 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && ( $PRECIS
             "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
             "configs/$SRTCTL_SETUP_SCRIPT"
     fi
+elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR" || exit 1
+    # Pin srt-slurm main so upstream movement does not change behavior
+    git checkout b959ff7d78c396f3f47bb88a115d8a492326aafe
+    mkdir -p recipes/vllm/kimi-k2.6-fp4
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4" recipes/vllm/kimi-k2.6-fp4
 else
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
@@ -184,6 +196,13 @@ SRTCTL_APPLY_ARGS=(
 if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
     SRTCTL_APPLY_ARGS+=(--setup-script "$SRTCTL_SETUP_SCRIPT")
 fi
+if [[ "$MODEL_PREFIX" == "kimik2.6" ]]; then
+    # MODEL_PATH lives on compute-node-local NVMe (/scratch/models), which is
+    # not mounted on the runner invoking srtctl, so skip the Python-level
+    # model-path preflight. vLLM still fails loudly at runtime if the path is
+    # genuinely missing on the compute node.
+    SRTCTL_APPLY_ARGS+=(--no-preflight)
+fi
 SRTCTL_OUTPUT=$(srtctl apply "${SRTCTL_APPLY_ARGS[@]}" 2>&1)
 echo "$SRTCTL_OUTPUT"
 

From 58e9bd6b48cc9016526fad8a4e199992a9b4eea4 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 13 Jul 2026 15:51:33 -0700
Subject: [PATCH 2/5] chore: set perf-changelog pr-link for #2181
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 perf-changelog 中的 pr-link 更新为 #2181。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f850b543e2..805165b7a3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4757,4 +4757,4 @@
     - "Add Kimi-K2.6 NVFP4 B300 disaggregated multinode vLLM benchmarks via Dynamo"
     - "Image: vllm/vllm-openai:v0.22.0"
     - "7 topologies at 8k/1k: 1p1d/3p1d/4p1d dep4-dep8 (4p1d also as FlashInfer-autotune variant at conc 4-4096), 1p8d dep4-tp4, 1p10d dep4-tp4 (FlashInfer autotune), 2p5d dep4-tp8; conc 1-4096; recipes under benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2181

From 3c325a9100f9b95bf9cc6c3b438be42ed19c87ed Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 13 Jul 2026 15:58:24 -0700
Subject: [PATCH 3/5] fix: add router and kv-p2p-transfer metadata to kimik2.6
 b300 disagg entry
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Main now requires kv-p2p-transfer on disagg entries (PR #2096); set nixl to
match the recipes' NixlConnector and add the dynamo-router metadata, matching
the other dynamo-vllm disagg entries.

中文：适配主分支新增的校验规则（PR #2096）：为 kimik2.6 B300 分离式配置补充
kv-p2p-transfer: nixl（与配方中的 NixlConnector 一致）及 dynamo-router 元数据，
与其他 dynamo-vllm 分离式配置保持一致。
---
 configs/nvidia-master.yaml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 50d3ee4d55..b32851f060 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -14178,6 +14178,8 @@ kimik2.6-fp4-b300-dynamo-vllm:
   runner: b300
   precision: fp4
   framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.2.0.dev20260529" }
+  kv-p2p-transfer: nixl
   multinode: true
   disagg: true
   scenarios:

From fde2506cd418dc2fbbc32fe32c690fce3072f507 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 13 Jul 2026 16:50:24 -0700
Subject: [PATCH 4/5] fix: drop hardcoded NixlConnector engine_id from kimik2.6
 b300 recipes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

A shared literal engine_id across independent prefill workers collides in
vLLM's decode-side NIXL handshake cache (keyed on engine_id alone), so
multi-prefill topologies could silently read KV from the wrong prefill
host. Omit the field so vLLM auto-generates a unique per-process id,
matching every other disagg recipe in the tree.

中文：移除 kimik2.6 B300 配方中硬编码的 NixlConnector engine_id。多个独立
预填充 worker 共用同一 engine_id 时，会在 vLLM 解码侧的 NIXL 握手缓存
（仅以 engine_id 为键）中发生冲突，多预填充拓扑可能静默地从错误的预填充
节点读取 KV。省略该字段后 vLLM 会为每个进程自动生成唯一 id，与仓库中其他
分离式配方保持一致。
---
 .../kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml    | 2 +-
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml     | 2 +-
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml      | 2 +-
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml      | 2 +-
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml     | 2 +-
 .../kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml    | 2 +-
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml     | 2 +-
 7 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
index d1ce703229..49ef80b401 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
index 7e2471d7a9..bdcaef3a06 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
index 965b118ffd..feb8da45be 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
index 7f5297121f..9f37a7ff39 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
index 847bef8744..9c1fa18a36 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
index edce630d10..0244b994b8 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
index 93401c50f9..a0da86982e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
@@ -39,7 +39,7 @@ backend:
     TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: "1800"
   vllm_config:
     prefill:
-      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both", "engine_id": "kimi-k26-prefill-dep4"}'
+      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

From 9567960513035e02413d2432cd36d47876e46644 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Mon, 13 Jul 2026 22:05:16 -0700
Subject: [PATCH 5/5] fix: use legacy CUDA_VISIBLE_DEVICES binding for kimik2.6
 b300 recipes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

srt-slurm at the pinned commit passes --device-ids to vLLM workers by
default, but that flag only exists in vLLM builds newer than v0.22.0
(vllm-project/vllm#45026, merged 2026-06-20), so every worker exited with
"unrecognized arguments". Set the schema's set_cuda_visible_devices
fallback so srtctl binds GPUs via CUDA_VISIBLE_DEVICES instead.

中文：srt-slurm 固定 commit 默认向 vLLM worker 传递 --device-ids，但该参数
仅存在于比 v0.22.0 更新的 vLLM 版本（vllm-project/vllm#45026，2026-06-20
合入），导致所有 worker 以 "unrecognized arguments" 退出。改用 schema 提供的
set_cuda_visible_devices 回退，使 srtctl 通过 CUDA_VISIBLE_DEVICES 绑定 GPU。
---
 .../kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml     | 1 +
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml      | 1 +
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml       | 1 +
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml       | 1 +
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml      | 1 +
 .../kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml     | 1 +
 .../vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml      | 1 +
 7 files changed, 7 insertions(+)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
index 49ef80b401..56e1d33493 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p10d-dep4-tp4-atune.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
index bdcaef3a06..88b96e38cb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
index feb8da45be..5be9475473 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-1p8d-dep4-tp4.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
index 9f37a7ff39..8f2efe52eb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-2p5d-dep4-tp8.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
index 9c1fa18a36..b77b6ca08b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-3p1d-dep4-dep8.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
index 0244b994b8..fc01cbcfee 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8-atune.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
index a0da86982e..5ade70c7ac 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6-fp4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
@@ -24,6 +24,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  set_cuda_visible_devices: true
   prefill_environment:
     VLLM_USE_FLASHINFER_MOE_FP4: "1"
     VLLM_USE_NCCL_SYMM_MEM: "1"