Merge remote-tracking branch 'origin/fix/single-node-dram-mem' into validate/refactor-stack

#3804 restores h100-dgxc to 2,063,837 MiB and main's H100 recipe values.
Keep the stack's '@dram' H100 recipes and its MiniMax-M3 H100
dram-utilization 0.4725 (the ~119 GiB Mooncake segment), so every H100
bound recipe matches refactor/recipe-bundles again.

Author: adibarra

inferencex-e2e/configs/CONFIGS.md

@@ -195,9 +195,9 @@ schema; unknown keys fail.
   master config's `cluster:<id>` label names; broad SKU labels use the GPU family when its
   clusters agree. Agentic master configs must use a `cluster:<id>` label; their DRAM
   KV-offload matrices combine both fields with `dram-utilization` into `total-cpu-dram-gb`.
-- `available-cpu-dram-mib` is the host DRAM, in MiB, a benchmark job can use on a node:
-  the lower of Slurm's `RealMemory - MemSpecLimit` and the node's `MemTotal` less 16 GiB for
-  the OS, over the cluster's GPU nodes. Omit it where unmeasured. When
+- `available-cpu-dram-mib` is the host DRAM, in MiB, a benchmark job can use on a node.
+  Measure it as the lower of Slurm's `RealMemory - MemSpecLimit` and the node's `MemTotal`
+  less 16 GiB for the OS, over the cluster's GPU nodes; omit it where unmeasured. When
   `slurm.srt-slurm.single-node-exclusive` is true (the default), a single-node srt-slurm job
   requests `--mem=0`, all the memory Slurm has on the node, and its recipe's host KV pools
   must fit the node. Otherwise the job requests, and its srun step is capped at, the serving

inferencex-e2e/configs/nvidia-master.yaml

@@ -5381,8 +5381,8 @@ minimaxm3-fp8-h100-vllm-agentic-mtp:
   scenarios:
     agentic-coding:
     # The fast sweep places the resident HBM cliff between c5 and c6.
-    # Sizes the 105 GiB Mooncake segment per GPU; weights and page cache use the rest.
-    - dram-utilization: 0.449
+    # Sizes the ~119 GiB Mooncake segment per GPU; weights and page cache use the rest.
+    - dram-utilization: 0.4725
       search-space:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 3, 4, 5], srt-recipe: agentic.yaml }
       - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [6, 8], srt-recipe: agentic.yaml }

inferencex-e2e/configs/runners.yaml

@@ -401,7 +401,7 @@ clusters:
         single-node-exclusive: false
   h100-dgxc:
     gpus-per-node: 8
-    available-cpu-dram-mib: 1_915_904
+    available-cpu-dram-mib: 2_063_837
     arch: x86_64
     models:
       entries: