From 1f066813518bdfdfbe8260a306ac6405c08c7f8d Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Wed, 30 Sep 2026 15:52:42 -0700
Subject: [PATCH 01/12] feat(config): enable MiniMax-M3 H100 FP8 indexer cache
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

启用 MiniMax-M3 H100 FP8 索引缓存，更新 vLLM 镜像并调整内存与 Mooncake 设置。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml     | 16 +++++++++-------
 inferencex-e2e/configs/nvidia-master.yaml        |  2 +-
 inferencex-e2e/perf-changelog.yaml               |  9 +++++++++
 3 files changed, 19 insertions(+), 8 deletions(-)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index dc2aa8f058..d212a5afb5 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -1,11 +1,11 @@
 # MiniMax-M3 MXFP8 AgentX on H100 with vLLM EAGLE3 (the GQA draft head) and
-# optional Mooncake DRAM KV offload. 26 GiB of weights per GPU live in host memory.
+# optional Mooncake DRAM KV offload and an FP8 MSA indexer cache.
 base:
   schema: 2
   name: minimaxm3-fp8-h100-vllm-agentic
   model:
     path: hf:MiniMaxAI/MiniMax-M3-MXFP8
-    container: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+    container: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
     precision: fp8
   resources:
     gpu_type: h100
@@ -35,11 +35,12 @@ base:
         served-model-name: MiniMaxAI/MiniMax-M3-MXFP8
         tensor-parallel-size: 8
         data-parallel-size: 1
-        gpu-memory-utilization: 0.90
-        cpu-offload-gb: 26
+        gpu-memory-utilization: 0.93
+        cpu-offload-gb: 5
         attention-backend: TRITON_ATTN
         safetensors-load-strategy: lazy
         kv-cache-dtype: fp8
+        attention-config: '{"indexer_kv_dtype":"fp8"}'
         block-size: 128
         language-model-only: true
         enable-prefix-caching: true
@@ -55,6 +56,7 @@ base:
       env:
         PYTHONNOUSERSITE: '1'
         VLLM_ENGINE_READY_TIMEOUT_S: '7200'
+        VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: '0'
   benchmark:
     type: custom
     command: bash /infmax-workspace/benchmarks/srt_agentic.sh
@@ -122,7 +124,7 @@ override_tp8_c5:
 
 # DRAM points offload KV to an embedded Mooncake store. Per rank: the host budget
 # (1731 GB = 1612 GiB) less the 414 GiB checkpoint page cache, over TP8, less the
-# 26 GiB CPU weight offload and the 4 GiB local buffer = 119 GB.
+# 5 GiB CPU weight offload and the 4 GiB local buffer = 140 GB.
 override_tp8_c6_dram:
   # The worker's Mooncake client and the master run the same pinned release.
   setup_script: vllm-mooncake-0.3.11.sh
@@ -144,7 +146,7 @@ override_tp8_c6_dram:
         store_config:
           mode: embedded
           metadata_server: P2PHANDSHAKE
-          global_segment_size: 119GB
+          global_segment_size: 140GB
           local_buffer_size: 4GB
           protocol: rdma
           device_name: ''
@@ -187,7 +189,7 @@ override_tp8_c8_dram:
         store_config:
           mode: embedded
           metadata_server: P2PHANDSHAKE
-          global_segment_size: 119GB
+          global_segment_size: 140GB
           local_buffer_size: 4GB
           protocol: rdma
           device_name: ''
diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index 63dcf8fff8..a009eb2b40 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -5367,7 +5367,7 @@ qwen3.5-fp4-b200-trt-mtp:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.5/trtllm/b200-fp4-mtp/8k1k.yaml
 
 minimaxm3-fp8-h100-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-ddd6fbca148a867aad1fcab7ec72f582b9977db4
+  image: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
   runner: cluster:h100-dgxc
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index b415cf56a5..0319275d67 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9158,3 +9158,12 @@
     - "Port the DeepSeek-V4-Pro-0813 MI355X ATOM 1P1D AgentX config from the legacy amd_utils path to a native srt-slurm recipe (benchmarks/multi_node/srt-slurm-recipes/dsv4/atom/mi355x-fp4/agentx/disagg-lmcache-dspark.yaml), one override variant per point. Server flags, env, AToMesh routing policies and decode CUDA-graph capture sizes match the legacy server_atom.sh for every tier: TP8 at concurrency 1 and 16, DP attention with prefill TBO at 64 and 128, and DP attention with ATOM's in-process LMCache CPU offload (lmcache_offload, 187 GB per prefill rank) at 256. Golden acceptance 3.01 for DSpark with three draft tokens is now injected by the srt-slurm path, the same value the legacy models_atom.yaml hardcoded."
     - "The LMCache tier uses srt-slurm's extra-kv-connectors (NVIDIA/srt-slurm#507, in v2.36.0) to add lmcache_offload next to the generated Mooncake connector."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3543
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the MiniMax-M3 H100 vLLM image, enable the FP8 MSA indexer cache, adjust CPU weight offload and GPU memory settings, and resize the Mooncake store segments."
+    - "更新 MiniMax-M3 H100 的 vLLM 镜像，启用 FP8 MSA 索引缓存，调整 CPU 权重卸载与 GPU 内存设置，并调整 Mooncake 存储段大小。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From 14a1e0081c0b5df8ac1bb29a80e7a77f6beffb35 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Wed, 30 Sep 2026 15:53:46 -0700
Subject: [PATCH 02/12] chore(config): link MiniMax-M3 H100 changelog entry
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

补充 MiniMax-M3 H100 变更日志条目的 PR 链接。
---
 inferencex-e2e/perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 0319275d67..2c6393edfe 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9166,4 +9166,4 @@
   description:
     - "Update the MiniMax-M3 H100 vLLM image, enable the FP8 MSA indexer cache, adjust CPU weight offload and GPU memory settings, and resize the Mooncake store segments."
     - "更新 MiniMax-M3 H100 的 vLLM 镜像，启用 FP8 MSA 索引缓存，调整 CPU 权重卸载与 GPU 内存设置，并调整 Mooncake 存储段大小。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 995dce32ef14f63000aa09b9a0be5c6b5fdc7c51 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Wed, 30 Sep 2026 23:55:14 -0700
Subject: [PATCH 03/12] fix(config): route MiniMax-M3 H100 sweep to DSXE
 runners
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 MiniMax-M3 H100 扫描任务切换到 DSXE 运行器，并添加对应的 Slurm 集群配置。
---
 .github/workflows/benchmark-tmpl.yml      |  2 ++
 inferencex-e2e/configs/nvidia-master.yaml |  2 +-
 inferencex-e2e/configs/runners.yaml       | 40 +++++++++++++++++++++++
 inferencex-e2e/perf-changelog.yaml        |  9 +++++
 4 files changed, 52 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/benchmark-tmpl.yml b/.github/workflows/benchmark-tmpl.yml
index 35548863fe..75c79fb14e 100644
--- a/.github/workflows/benchmark-tmpl.yml
+++ b/.github/workflows/benchmark-tmpl.yml
@@ -164,8 +164,10 @@ permissions:
 
 jobs:
   benchmark:
+    # DSXE runners currently advertise h100-dgxc-new without a cluster label.
     runs-on: >-
       ${{ fromJSON(
+        inputs.runner == 'cluster:h100-dsxe' && '["self-hosted","h100-dgxc-new"]' ||
         vars.PRIORITY_SCHEDULER_ENABLED == 'true' &&
         (
           vars.NODE_SLOT_SCHEDULER_ENABLED == 'true' &&
diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index baadd0ccd1..a0916b9ed1 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -5370,7 +5370,7 @@ minimaxm3-fp8-h100-vllm-agentic-mtp:
   image: vllm/vllm-openai:nightly-36768d1bfd39094681cdbc8cb37d4b31c0729c89
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
-  runner: cluster:h100-cw
+  runner: cluster:h100-dsxe
   precision: fp8
   framework: vllm
   multinode: false
diff --git a/inferencex-e2e/configs/runners.yaml b/inferencex-e2e/configs/runners.yaml
index 8036d3de69..b4f2773b2b 100644
--- a/inferencex-e2e/configs/runners.yaml
+++ b/inferencex-e2e/configs/runners.yaml
@@ -2,6 +2,9 @@ labels:
   h100:
     - h100-cw_00
     - h100-cw_01
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
     - h100-dgxc-slurm_00
     - h100-dgxc-slurm_01
     - h100-dgxc-slurm_02
@@ -156,6 +159,14 @@ labels:
   cluster:h100-cw:
     - h100-cw_00
     - h100-cw_01
+  h100-dgxc-new:
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
+  cluster:h100-dsxe:
+    - h100-dgxc-new_00
+    - h100-dgxc-new_01
+    - h100-dgxc-new_02
   cluster:h100-dgxc:
     - h100-dgxc-slurm_00
     - h100-dgxc-slurm_01
@@ -326,6 +337,35 @@ clusters:
         network-interface: ""
         gpus-per-node-directive: false
         single-node-exclusive: false
+  h100-dsxe:
+    gpus-per-node: 8
+    available-cpu-dram-mib: 1_998_848
+    arch: x86_64
+    models:
+      entries:
+        MiniMax-M3-MXFP8: {root: data-models, dir: MiniMax-M3-MXFP8}
+    scheduler: slurm
+    slurm:
+      partition: batch_1
+      account: benchmark
+      exclusive: false
+      gres: "gpu:{gpus}"
+      volumes:
+        hf-hub-cache: {path: /data/home/sa-gha-runner/gharunners/hf-hub-cache}
+        aiperf-cache: {path: /data/home/sa-gha-runner/gharunners/ai-perf-cache}
+        data-models: {path: /data/models}
+      squash:
+        dir: /data/home/sa-gha-runner/gharunners/squash
+        visibility: shared
+        import: compute
+        lock-timeout-s: 600
+      srt-slurm:
+        extra:
+          default_gpu_exporter:
+            container_image: nvcr.io#nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless
+            port: 9401
+            command: "dcgm-exporter --collect-interval=1000 --address :{port} -f /configs/dcgm-counters-noprof.csv"
+        network-interface: ""
   h100-dgxc:
     gpus-per-node: 8
     available-cpu-dram-mib: 2_063_837
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 7cbd3d336c..ae84c99347 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9186,3 +9186,12 @@
     - "Route the MiniMax-M3 H100 AgentX config to cluster:h100-cw and align the Mooncake DRAM store size with that cluster's host-memory budget."
     - "将 MiniMax-M3 H100 AgentX 配置切换到 cluster:h100-cw，并根据该集群的主机内存预算调整 Mooncake DRAM 存储段大小。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Route the MiniMax-M3 H100 AgentX config to the DSXE runner group and add its Slurm cluster profile."
+    - "将 MiniMax-M3 H100 AgentX 配置切换到 DSXE 运行器组，并添加对应的 Slurm 集群配置。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From f8ef6a7d65d4a2d2a9258e2e667ada239b2e0c85 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 1 Oct 2026 00:02:12 -0700
Subject: [PATCH 04/12] fix(config): use staged MiniMax-M3 checkpoint on DSXE
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

让 DSXE 上的 MiniMax-M3 单节点任务读取已暂存的模型权重。
---
 inferencex-e2e/configs/runners.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/inferencex-e2e/configs/runners.yaml b/inferencex-e2e/configs/runners.yaml
index b4f2773b2b..96da78e36a 100644
--- a/inferencex-e2e/configs/runners.yaml
+++ b/inferencex-e2e/configs/runners.yaml
@@ -366,6 +366,7 @@ clusters:
             port: 9401
             command: "dcgm-exporter --collect-interval=1000 --address :{port} -f /configs/dcgm-counters-noprof.csv"
         network-interface: ""
+        single-node-models: staged
   h100-dgxc:
     gpus-per-node: 8
     available-cpu-dram-mib: 2_063_837

From 051c37f18f0a72b8b879d0e268b1ad1bd58db9a4 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 1 Oct 2026 05:41:25 -0700
Subject: [PATCH 05/12] fix(config): use TCP for MiniMax-M3 Mooncake DRAM store
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 MiniMax-M3 的 Mooncake DRAM 存储切换为 TCP 传输，避免在单节点上注册大型 RDMA 内存段。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml             | 4 ++--
 inferencex-e2e/perf-changelog.yaml                       | 9 +++++++++
 2 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 0a947e7c9a..8411f65384 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -146,7 +146,7 @@ override_tp8_c6_dram:
           metadata_server: P2PHANDSHAKE
           global_segment_size: 134GB
           local_buffer_size: 4GB
-          protocol: rdma
+          protocol: tcp
           device_name: ''
           enable_offload: false
   roles:
@@ -189,7 +189,7 @@ override_tp8_c8_dram:
           metadata_server: P2PHANDSHAKE
           global_segment_size: 134GB
           local_buffer_size: 4GB
-          protocol: rdma
+          protocol: tcp
           device_name: ''
           enable_offload: false
   roles:
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index ae84c99347..5c03d88d02 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9195,3 +9195,12 @@
     - "Route the MiniMax-M3 H100 AgentX config to the DSXE runner group and add its Slurm cluster profile."
     - "将 MiniMax-M3 H100 AgentX 配置切换到 DSXE 运行器组，并添加对应的 Slurm 集群配置。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use TCP transport for the single-node Mooncake DRAM store in the concurrency 6 and 8 variants."
+    - "将并发数为 6 和 8 的单节点 Mooncake DRAM 存储变体切换为 TCP 传输。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 0e36fba0a1e5224b06bf43d7ad4f27ba40af701b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 1 Oct 2026 10:43:18 -0500
Subject: [PATCH 06/12] fix(ci): use cluster labels and priority leases for
 H100 DSXE

---
 .github/workflows/benchmark-tmpl.yml | 2 --
 1 file changed, 2 deletions(-)

diff --git a/.github/workflows/benchmark-tmpl.yml b/.github/workflows/benchmark-tmpl.yml
index 75c79fb14e..35548863fe 100644
--- a/.github/workflows/benchmark-tmpl.yml
+++ b/.github/workflows/benchmark-tmpl.yml
@@ -164,10 +164,8 @@ permissions:
 
 jobs:
   benchmark:
-    # DSXE runners currently advertise h100-dgxc-new without a cluster label.
     runs-on: >-
       ${{ fromJSON(
-        inputs.runner == 'cluster:h100-dsxe' && '["self-hosted","h100-dgxc-new"]' ||
         vars.PRIORITY_SCHEDULER_ENABLED == 'true' &&
         (
           vars.NODE_SLOT_SCHEDULER_ENABLED == 'true' &&

From f1050134ef61888961cb2ddec357fd7419cf5b4a Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 1 Oct 2026 11:42:53 -0700
Subject: [PATCH 07/12] fix(config): reserve GPU memory for Mooncake TCP
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

为 Mooncake TCP 传输预留 GPU 内存。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml             | 4 ++++
 inferencex-e2e/perf-changelog.yaml                       | 9 +++++++++
 2 files changed, 13 insertions(+)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 8411f65384..fe826f4336 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -152,6 +152,8 @@ override_tp8_c6_dram:
   roles:
     agg:
       args:
+        # Leave room for Mooncake TCP CUDA contexts on GPU 0 during KV transfers.
+        gpu-memory-utilization: 0.85
         max-num-seqs: 12
         max-cudagraph-capture-size: 48
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
@@ -195,6 +197,8 @@ override_tp8_c8_dram:
   roles:
     agg:
       args:
+        # Leave room for Mooncake TCP CUDA contexts on GPU 0 during KV transfers.
+        gpu-memory-utilization: 0.85
         max-num-seqs: 16
         max-cudagraph-capture-size: 64
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index c1cb090e40..1b8e1e50c8 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9216,3 +9216,12 @@
     - "Use TCP transport for the single-node Mooncake DRAM store in the concurrency 6 and 8 variants."
     - "将并发数为 6 和 8 的单节点 Mooncake DRAM 存储变体切换为 TCP 传输。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Reserve GPU memory for Mooncake TCP transfers in the MiniMax-M3 H100 DRAM variants."
+    - "为 MiniMax-M3 H100 DRAM 变体的 Mooncake TCP 传输预留 GPU 内存。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 4467aab4f78b07b23d4097cc457c4ecf23a3bd8c Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 1 Oct 2026 16:25:07 -0700
Subject: [PATCH 08/12] fix(config): align MiniMax DRAM context limits
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

为 MiniMax DRAM 变体统一服务与重放上下文上限。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml             | 4 ++++
 inferencex-e2e/perf-changelog.yaml                       | 9 +++++++++
 2 files changed, 13 insertions(+)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index fe826f4336..835916b076 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -154,6 +154,7 @@ override_tp8_c6_dram:
       args:
         # Leave room for Mooncake TCP CUDA contexts on GPU 0 during KV transfers.
         gpu-memory-utilization: 0.85
+        max-model-len: 655360
         max-num-seqs: 12
         max-cudagraph-capture-size: 48
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
@@ -167,6 +168,7 @@ override_tp8_c6_dram:
       CONC: '6'
       KV_OFFLOADING: dram
       TOTAL_CPU_DRAM_GB: '1676'
+      AIPERF_MAX_CONTEXT_LENGTH: '655360'
 
 override_tp8_c8_dram:
   # The worker's Mooncake client and the master run the same pinned release.
@@ -199,6 +201,7 @@ override_tp8_c8_dram:
       args:
         # Leave room for Mooncake TCP CUDA contexts on GPU 0 during KV transfers.
         gpu-memory-utilization: 0.85
+        max-model-len: 655360
         max-num-seqs: 16
         max-cudagraph-capture-size: 64
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
@@ -212,3 +215,4 @@ override_tp8_c8_dram:
       CONC: '8'
       KV_OFFLOADING: dram
       TOTAL_CPU_DRAM_GB: '1676'
+      AIPERF_MAX_CONTEXT_LENGTH: '655360'
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 1b8e1e50c8..99ca7b9db7 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9225,3 +9225,12 @@
     - "Reserve GPU memory for Mooncake TCP transfers in the MiniMax-M3 H100 DRAM variants."
     - "为 MiniMax-M3 H100 DRAM 变体的 Mooncake TCP 传输预留 GPU 内存。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Set matching 640K service and replay context limits for the MiniMax-M3 H100 Mooncake DRAM variants."
+    - "为 MiniMax-M3 H100 Mooncake DRAM 变体设置一致的 640K 服务与重放上下文上限。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 439616c1bbab5d1c626df5bf5b83dbe7aeaac8b3 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Thu, 1 Oct 2026 19:25:16 -0700
Subject: [PATCH 09/12] fix(config): reuse MiniMax Mooncake TCP connections
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

在 MiniMax Mooncake DRAM 变体中复用 TCP 连接，避免短连接重复创建。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml             | 2 ++
 inferencex-e2e/perf-changelog.yaml                       | 9 +++++++++
 2 files changed, 11 insertions(+)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 835916b076..0110f2a780 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -162,6 +162,7 @@ override_tp8_c6_dram:
         PYTHONHASHSEED: '0'
         MC_SLICE_SIZE: '1048576'
         MC_WORKERS_PER_CTX: '4'
+        MC_TCP_ENABLE_CONNECTION_POOL: '1'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
@@ -209,6 +210,7 @@ override_tp8_c8_dram:
         PYTHONHASHSEED: '0'
         MC_SLICE_SIZE: '1048576'
         MC_WORKERS_PER_CTX: '4'
+        MC_TCP_ENABLE_CONNECTION_POOL: '1'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 4c00cefcfa..3b9a4b3112 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9256,3 +9256,12 @@
     - "Set matching 640K service and replay context limits for the MiniMax-M3 H100 Mooncake DRAM variants."
     - "为 MiniMax-M3 H100 Mooncake DRAM 变体设置一致的 640K 服务与重放上下文上限。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Reuse Mooncake TCP connections in the MiniMax-M3 H100 DRAM variants."
+    - "在 MiniMax-M3 H100 DRAM 变体中复用 Mooncake TCP 连接。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 5f09e094577842a6925c578df7785293bab65de8 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 2 Oct 2026 01:02:08 -0700
Subject: [PATCH 10/12] fix: increase Mooncake transfer slices for H100 DRAM
 runs
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Set larger TCP transfer slices for the MiniMax-M3 H100 DRAM variants and use a separate live replay error threshold. Keep final result validation unchanged.

中文：增大 MiniMax-M3 H100 DRAM 变体的 Mooncake TCP 传输分片，并单独设置重放过程中的实时错误阈值。最终结果校验保持不变。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml             | 6 ++++--
 inferencex-e2e/perf-changelog.yaml                       | 9 +++++++++
 2 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 0110f2a780..6841db940b 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -160,7 +160,7 @@ override_tp8_c6_dram:
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
       env:
         PYTHONHASHSEED: '0'
-        MC_SLICE_SIZE: '1048576'
+        MC_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
@@ -170,6 +170,7 @@ override_tp8_c6_dram:
       KV_OFFLOADING: dram
       TOTAL_CPU_DRAM_GB: '1676'
       AIPERF_MAX_CONTEXT_LENGTH: '655360'
+      AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: '0.25'
 
 override_tp8_c8_dram:
   # The worker's Mooncake client and the master run the same pinned release.
@@ -208,7 +209,7 @@ override_tp8_c8_dram:
         kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
       env:
         PYTHONHASHSEED: '0'
-        MC_SLICE_SIZE: '1048576'
+        MC_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
@@ -218,3 +219,4 @@ override_tp8_c8_dram:
       KV_OFFLOADING: dram
       TOTAL_CPU_DRAM_GB: '1676'
       AIPERF_MAX_CONTEXT_LENGTH: '655360'
+      AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: '0.25'
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 64b04b8944..f5a1f42362 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9274,3 +9274,12 @@
     - "Reuse Mooncake TCP connections in the MiniMax-M3 H100 DRAM variants."
     - "在 MiniMax-M3 H100 DRAM 变体中复用 Mooncake TCP 连接。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use larger Mooncake transfer slices and a separate live replay error threshold for the MiniMax-M3 H100 DRAM variants; retain the final result validation threshold."
+    - "为 MiniMax-M3 H100 DRAM 变体使用更大的 Mooncake 传输分片，并单独设置重放过程中的实时错误阈值；最终结果校验阈值保持不变。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 1bebab03cf231cab3b1f370c78bcd343934eb0d4 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 2 Oct 2026 04:27:16 -0700
Subject: [PATCH 11/12] fix: bound Mooncake TCP lanes for H100 DRAM variants
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

为 H100 DRAM 变体使用有上限的 Mooncake TCP 连接通道。
---
 .../srt-slurm-recipes/configs/vllm-mooncake-0.3.13.sh  |  9 +++++++++
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml           | 10 ++++++----
 inferencex-e2e/perf-changelog.yaml                     |  9 +++++++++
 3 files changed, 24 insertions(+), 4 deletions(-)
 create mode 100755 inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-mooncake-0.3.13.sh

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-mooncake-0.3.13.sh b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-mooncake-0.3.13.sh
new file mode 100755
index 0000000000..b4301604da
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/vllm-mooncake-0.3.13.sh
@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# Pin the worker's Mooncake store client to the release the master runs.
+set -eo pipefail
+pip_install=(python3 -m pip install)
+if python3 -m pip install --help 2>/dev/null | grep -q -- --break-system-packages; then
+    pip_install+=(--break-system-packages)
+fi
+"${pip_install[@]}" --quiet --no-cache-dir --no-deps --force-reinstall mooncake-transfer-engine-cuda13==0.3.13.post1
+python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 6841db940b..8b2eaba115 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -125,7 +125,7 @@ override_tp8_c5:
 # 5 GiB CPU weight offload and the 4 GiB local buffer = 134 GB.
 override_tp8_c6_dram:
   # The worker's Mooncake client and the master run the same pinned release.
-  setup_script: vllm-mooncake-0.3.11.sh
+  setup_script: vllm-mooncake-0.3.13.sh
   services:
     - name: mooncake-master
       type: mooncake-master
@@ -136,7 +136,7 @@ override_tp8_c6_dram:
       preamble: >-
         pip_install=(python3 -m pip install);
         python3 -m pip install --help 2>/dev/null | grep -q -- --break-system-packages && pip_install+=(--break-system-packages);
-        "${pip_install[@]}" --no-cache-dir --no-deps --force-reinstall mooncake-transfer-engine-cuda13==0.3.11.post1
+        "${pip_install[@]}" --no-cache-dir --no-deps --force-reinstall mooncake-transfer-engine-cuda13==0.3.13.post1
       args:
         - '--eviction_high_watermark_ratio=0.80'
         - '--eviction_ratio=0.10'
@@ -163,6 +163,7 @@ override_tp8_c6_dram:
         MC_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
+        MC_TCP_LANES_PER_PEER: '4'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
@@ -174,7 +175,7 @@ override_tp8_c6_dram:
 
 override_tp8_c8_dram:
   # The worker's Mooncake client and the master run the same pinned release.
-  setup_script: vllm-mooncake-0.3.11.sh
+  setup_script: vllm-mooncake-0.3.13.sh
   services:
     - name: mooncake-master
       type: mooncake-master
@@ -185,7 +186,7 @@ override_tp8_c8_dram:
       preamble: >-
         pip_install=(python3 -m pip install);
         python3 -m pip install --help 2>/dev/null | grep -q -- --break-system-packages && pip_install+=(--break-system-packages);
-        "${pip_install[@]}" --no-cache-dir --no-deps --force-reinstall mooncake-transfer-engine-cuda13==0.3.11.post1
+        "${pip_install[@]}" --no-cache-dir --no-deps --force-reinstall mooncake-transfer-engine-cuda13==0.3.13.post1
       args:
         - '--eviction_high_watermark_ratio=0.80'
         - '--eviction_ratio=0.10'
@@ -212,6 +213,7 @@ override_tp8_c8_dram:
         MC_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
+        MC_TCP_LANES_PER_PEER: '4'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index f5a1f42362..fd7f462de9 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9283,3 +9283,12 @@
     - "Use larger Mooncake transfer slices and a separate live replay error threshold for the MiniMax-M3 H100 DRAM variants; retain the final result validation threshold."
     - "为 MiniMax-M3 H100 DRAM 变体使用更大的 Mooncake 传输分片，并单独设置重放过程中的实时错误阈值；最终结果校验阈值保持不变。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Pin the H100 DRAM variants to Mooncake 0.3.13.post1 with bounded TCP connection lanes."
+    - "将 H100 DRAM 变体固定为 Mooncake 0.3.13.post1，并使用有上限的 TCP 连接通道。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620

From 0cfb6ac1db447de975df53407a797bc7d703ea17 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Fri, 2 Oct 2026 07:54:05 -0700
Subject: [PATCH 12/12] fix: tune Mooncake TCP lane admission for H100 DRAM
 variants
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

调整 H100 DRAM 变体的 Mooncake TCP 连接通道接纳参数。
---
 .../minimaxm3/vllm/h100-fp8-mtp/agentic.yaml         | 12 ++++++++++--
 inferencex-e2e/perf-changelog.yaml                   |  9 +++++++++
 2 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
index 8b2eaba115..f4e8d21986 100644
--- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
+++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/h100-fp8-mtp/agentic.yaml
@@ -161,9 +161,13 @@ override_tp8_c6_dram:
       env:
         PYTHONHASHSEED: '0'
         MC_SLICE_SIZE: '8388608'
+        MC_TCP_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
-        MC_TCP_LANES_PER_PEER: '4'
+        MC_TCP_LANES_PER_PEER: '16'
+        MC_TCP_MAX_QUEUED_TRANSFERS_PER_PEER: '65535'
+        MC_TCP_MAX_PENDING_ADMISSIONS_PER_PEER: '65535'
+        MC_TCP_ADMISSION_TIMEOUT_MS: '60000'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
@@ -211,9 +215,13 @@ override_tp8_c8_dram:
       env:
         PYTHONHASHSEED: '0'
         MC_SLICE_SIZE: '8388608'
+        MC_TCP_SLICE_SIZE: '8388608'
         MC_WORKERS_PER_CTX: '4'
         MC_TCP_ENABLE_CONNECTION_POOL: '1'
-        MC_TCP_LANES_PER_PEER: '4'
+        MC_TCP_LANES_PER_PEER: '16'
+        MC_TCP_MAX_QUEUED_TRANSFERS_PER_PEER: '65535'
+        MC_TCP_MAX_PENDING_ADMISSIONS_PER_PEER: '65535'
+        MC_TCP_ADMISSION_TIMEOUT_MS: '60000'
         MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
   benchmark:
     env:
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index fd7f462de9..21e1581558 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9292,3 +9292,12 @@
     - "Pin the H100 DRAM variants to Mooncake 0.3.13.post1 with bounded TCP connection lanes."
     - "将 H100 DRAM 变体固定为 Mooncake 0.3.13.post1，并使用有上限的 TCP 连接通道。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Set TCP transfer slice size and bounded lane admission settings for the MiniMax-M3 H100 DRAM variants."
+    - "为 MiniMax-M3 H100 DRAM 变体设置 TCP 传输分片大小和有上限的连接通道接纳参数。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3620