Migrate Kimi-K3 AgentX recipe benchmark placement for srt-slurm v2.43.4 / 迁移 Kimi-K3 AgentX 配置的 benchmark placement 以适配 srt-slurm v2.43.4

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/RECIPES.md

@@ -4,7 +4,7 @@
 
 InferenceX owns the recipes in this directory. For every NVIDIA srt-slurm launch, the srt driver ([`infx/launch/drivers/srt/checkout.py`](../../../infx/launch/drivers/srt/checkout.py)) makes a job-local Git clone of the pinned submodule and copies this entire tree into `recipes/`. It records the actual revision in `srt-slurm-sha.txt`; power lanes copy that revision into `power-producer-sha.txt` for result validation.
 
-The shared version is the Git submodule pointer at [`utils/srt-slurm`](../../../utils/srt-slurm), currently [v2.36.0](https://github.com/NVIDIA/srt-slurm/releases/tag/v2.36.0) (`7b5863a7837673d81403b076be219bbf18a7700f`). Update that submodule pointer when upgrading, then run the recipe and integration checks. Do not add model-specific checkout branches to launchers.
+The shared version is the Git submodule pointer at [`utils/srt-slurm`](../../../utils/srt-slurm), currently [v2.43.4](https://github.com/NVIDIA/srt-slurm/releases/tag/v2.43.4) (`848f72d45b05af0fc082a9d1df49a8b4e7e61507`). Update that submodule pointer when upgrading, then run the recipe and integration checks. Do not add model-specific checkout branches to launchers.
 
 InferenceX requires srt-slurm 2.0 or newer and `schema: 2` recipes. Legacy recipe layouts are unsupported; migrate them before adding them to this tree.
 
@@ -54,8 +54,6 @@ All referenced recipes must be checked in: srt-slurm 2 ships curated examples in
 Install the shared pin in an isolated environment, then use its CLI:
 
 ```bash
-# Verify each supported recipe directory before rewriting it.
-srtctl migrate --verify -f benchmarks/multi_node/srt-slurm-recipes/dsr1/sglang
 srtctl migrate --in-place -f benchmarks/multi_node/srt-slurm-recipes/dsr1/sglang
 # Repeat for the other model/engine directories.
 python -m pytest infx/tests/matrix/ -q
@@ -64,7 +62,7 @@ python -m infx.matrix.generate full-sweep \
   --framework dynamo-sglang dynamo-trt dynamo-vllm --multi-node
 ```
 
-Validate recipes with the exact launcher pin, including all override variants. For a path-only reorganization, compare generated matrices before and after with the path mapping applied; all other fields, including eval selection and node counts, must match. A passing local schema check does not replace the full hardware sweep and evals.
+Validate recipes with the exact launcher pin, including all override variants. The current migration CLI has no `--verify`; use `srtctl dry-run` for each selected recipe after migration, supplying launcher-provided values such as `benchmark.concurrencies` through `--set` when needed. For a path-only reorganization, compare generated matrices before and after with the path mapping applied; all other fields, including eval selection and node counts, must match. A passing local schema check does not replace the full hardware sweep and evals.
 
 The initial migration also resolves compatibility issues that `srtctl migrate` cannot fix itself:
 

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/RECIPES_zh.md

@@ -4,7 +4,7 @@
 
 InferenceX 负责维护本目录中的配置。每次 NVIDIA srt-slurm 启动时，srt 驱动（[`infx/launch/drivers/srt/checkout.py`](../../../infx/launch/drivers/srt/checkout.py)）都会为作业创建固定版本子模块的本地 Git 克隆，并将整个目录复制到 `recipes/`。它将实际提交记录到 `srt-slurm-sha.txt`；功耗测试路径还会将其复制到 `power-producer-sha.txt`，供结果校验使用。
 
-统一版本由 [`utils/srt-slurm`](../../../utils/srt-slurm) 的 Git 子模块指针指定，目前为 [v2.36.0](https://github.com/NVIDIA/srt-slurm/releases/tag/v2.36.0)（`7b5863a7837673d81403b076be219bbf18a7700f`）。升级时更新该子模块指针，然后运行配置和集成检查。不要在启动器中新增按模型选择检出版本的分支。
+统一版本由 [`utils/srt-slurm`](../../../utils/srt-slurm) 的 Git 子模块指针指定，目前为 [v2.43.4](https://github.com/NVIDIA/srt-slurm/releases/tag/v2.43.4)（`848f72d45b05af0fc082a9d1df49a8b4e7e61507`）。升级时更新该子模块指针，然后运行配置和集成检查。不要在启动器中新增按模型选择检出版本的分支。
 
 InferenceX 要求 srt-slurm 2.0 或更新版本，且配置必须声明 `schema: 2`。不支持旧版配置结构；加入本目录前必须先完成迁移。
 
@@ -54,8 +54,6 @@ TileRT 使用固定版本的上游 srt-slurm 子模块。配置指定 `roles.pre
 在隔离环境中安装统一版本，然后使用其 CLI：
 
 ```bash
-# 重写前先验证每个受支持的配置目录。
-srtctl migrate --verify -f benchmarks/multi_node/srt-slurm-recipes/dsr1/sglang
 srtctl migrate --in-place -f benchmarks/multi_node/srt-slurm-recipes/dsr1/sglang
 # 对其他模型/引擎目录重复执行。
 python -m pytest infx/tests/matrix/ -q
@@ -64,7 +62,7 @@ python -m infx.matrix.generate full-sweep \
   --framework dynamo-sglang dynamo-trt dynamo-vllm --multi-node
 ```
 
-使用启动器指定的确切提交验证配置，包括全部覆盖变体。仅调整路径时，应按路径映射比较变更前后的生成矩阵；其他字段（包括评估选择和节点数）必须完全一致。本地配置校验通过不能替代完整硬件扫描和准确性评估。
+使用启动器指定的确切提交验证配置，包括全部覆盖变体。当前迁移 CLI 已不提供 `--verify`；迁移后对每个选中的配置运行 `srtctl dry-run`，必要时通过 `--set` 提供启动器注入的值，例如 `benchmark.concurrencies`。仅调整路径时，应按路径映射比较变更前后的生成矩阵；其他字段（包括评估选择和节点数）必须完全一致。本地配置校验通过不能替代完整硬件扫描和准确性评估。
 
 本次迁移还修复了 `srtctl migrate` 无法自动处理的兼容性问题：
 

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/mi355x-fp4/agentx/disagg-umbp-dspark.yaml

@@ -2,6 +2,7 @@
 # with DSpark. Every arm offloads prefill KV to UMBP as a direct external store
 # for the unified radix tree (no host cache tier), backed by a 1.5 TB
 # hugepage DRAM tier on the prefill node.
+schema: 2
 base:
   name: mi355x-dsv4-pro-0813-agentx-umbp
   model:

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/h200-fp8/agentx/agg-h200-tp8-mtp-c12.yaml

@@ -1,3 +1,4 @@
+schema: 2
 name: agg-h200-tp8-mtp-c12
 
 # Aggregated TP8 GLM-5.2 AgentX topology on one 8-GPU H200 node. MTP uses the
@@ -27,14 +28,17 @@ dynamo:
 resources:
   gpu_type: h200
   gpus_per_node: 8
-  gpus_per_agg: 8
-  agg_nodes: 1
-  agg_workers: 1
-
-infra:
-  etcd_nats_dedicated_node: false
-  nats_max_payload_mb: 64
-
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 64
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -49,10 +53,27 @@ frontend:
     active-prefill-tokens-threshold: None
     active-prefill-tokens-threshold-frac: None
 
-backend:
-  type: sglang
-  sglang_config:
-    aggregated:
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+
+    gpus: 8
+    env:
+      HF_HOME: /hf_hub_cache
+      HF_HUB_CACHE: /hf_hub_cache
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      PYTHONUNBUFFERED: "1"
+      SGLANG_OPT_USE_TOPK_V2: "1"
+      SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+      SGLANG_ENABLE_THINKING: "1"
+      SGLANG_REASONING_EFFORT: max
+      SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
+      SGLANG_HICACHE_DEBUG_LOG: "1"
+      SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
+
+    args:
       served-model-name: zai-org/GLM-5.2-FP8
       trust-remote-code: true
       tool-call-parser: glm47
@@ -83,22 +104,6 @@ backend:
       stream-interval: 60
       enable-metrics: true
       enable-cache-report: true
-  aggregated_environment:
-    HF_HOME: /hf_hub_cache
-    HF_HUB_CACHE: /hf_hub_cache
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-    PYTHONUNBUFFERED: "1"
-    SGLANG_OPT_USE_TOPK_V2: "1"
-    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
-    SGLANG_ENABLE_THINKING: "1"
-    SGLANG_REASONING_EFFORT: max
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
-    SGLANG_HICACHE_DEBUG_LOG: "1"
-    SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
-    SGLANG_SIMULATE_ACC_LEN: "2.99"
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: real-draft-token
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -118,7 +123,6 @@ telemetry:
 
 benchmark:
   type: custom
-  use_chat_template: true
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/h200-fp8/agentx/agg-h200-tp8-mtp-c16.yaml

@@ -1,3 +1,4 @@
+schema: 2
 name: agg-h200-tp8-mtp-c16
 
 # Aggregated TP8 GLM-5.2 AgentX topology on one 8-GPU H200 node. MTP uses the
@@ -27,14 +28,17 @@ dynamo:
 resources:
   gpu_type: h200
   gpus_per_node: 8
-  gpus_per_agg: 8
-  agg_nodes: 1
-  agg_workers: 1
-
-infra:
-  etcd_nats_dedicated_node: false
-  nats_max_payload_mb: 64
-
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 64
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -49,10 +53,27 @@ frontend:
     active-prefill-tokens-threshold: None
     active-prefill-tokens-threshold-frac: None
 
-backend:
-  type: sglang
-  sglang_config:
-    aggregated:
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+
+    gpus: 8
+    env:
+      HF_HOME: /hf_hub_cache
+      HF_HUB_CACHE: /hf_hub_cache
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      PYTHONUNBUFFERED: "1"
+      SGLANG_OPT_USE_TOPK_V2: "1"
+      SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+      SGLANG_ENABLE_THINKING: "1"
+      SGLANG_REASONING_EFFORT: max
+      SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
+      SGLANG_HICACHE_DEBUG_LOG: "1"
+      SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
+
+    args:
       served-model-name: zai-org/GLM-5.2-FP8
       trust-remote-code: true
       tool-call-parser: glm47
@@ -83,22 +104,6 @@ backend:
       stream-interval: 60
       enable-metrics: true
       enable-cache-report: true
-  aggregated_environment:
-    HF_HOME: /hf_hub_cache
-    HF_HUB_CACHE: /hf_hub_cache
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-    PYTHONUNBUFFERED: "1"
-    SGLANG_OPT_USE_TOPK_V2: "1"
-    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
-    SGLANG_ENABLE_THINKING: "1"
-    SGLANG_REASONING_EFFORT: max
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
-    SGLANG_HICACHE_DEBUG_LOG: "1"
-    SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
-    SGLANG_SIMULATE_ACC_LEN: "2.99"
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: real-draft-token
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -118,7 +123,6 @@ telemetry:
 
 benchmark:
   type: custom
-  use_chat_template: true
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/h200-fp8/agentx/agg-h200-tp8-mtp-c2.yaml

@@ -1,3 +1,4 @@
+schema: 2
 name: agg-h200-tp8-mtp-c2
 
 # Aggregated TP8 GLM-5.2 AgentX topology on one 8-GPU H200 node. MTP uses the
@@ -27,14 +28,17 @@ dynamo:
 resources:
   gpu_type: h200
   gpus_per_node: 8
-  gpus_per_agg: 8
-  agg_nodes: 1
-  agg_workers: 1
-
-infra:
-  etcd_nats_dedicated_node: false
-  nats_max_payload_mb: 64
-
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 64
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -49,10 +53,27 @@ frontend:
     active-prefill-tokens-threshold: None
     active-prefill-tokens-threshold-frac: None
 
-backend:
-  type: sglang
-  sglang_config:
-    aggregated:
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+
+    gpus: 8
+    env:
+      HF_HOME: /hf_hub_cache
+      HF_HUB_CACHE: /hf_hub_cache
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      PYTHONUNBUFFERED: "1"
+      SGLANG_OPT_USE_TOPK_V2: "1"
+      SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+      SGLANG_ENABLE_THINKING: "1"
+      SGLANG_REASONING_EFFORT: max
+      SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
+      SGLANG_HICACHE_DEBUG_LOG: "1"
+      SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
+
+    args:
       served-model-name: zai-org/GLM-5.2-FP8
       trust-remote-code: true
       tool-call-parser: glm47
@@ -83,22 +104,6 @@ backend:
       stream-interval: 60
       enable-metrics: true
       enable-cache-report: true
-  aggregated_environment:
-    HF_HOME: /hf_hub_cache
-    HF_HUB_CACHE: /hf_hub_cache
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-    PYTHONUNBUFFERED: "1"
-    SGLANG_OPT_USE_TOPK_V2: "1"
-    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
-    SGLANG_ENABLE_THINKING: "1"
-    SGLANG_REASONING_EFFORT: max
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
-    SGLANG_HICACHE_DEBUG_LOG: "1"
-    SGLANG_HICACHE_DEBUG_SAMPLE_RATE: "16384"
-    SGLANG_SIMULATE_ACC_LEN: "2.99"
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: real-draft-token
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -118,7 +123,6 @@ telemetry:
 
 benchmark:
   type: custom
-  use_chat_template: true
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace