feat(runners): add the B300 DSXE cluster and retire the B300 NV launcher / 添加 B300 DSXE 集群并下线 B300 NV 启动脚本

.github/workflows/benchmark-tmpl.yml

@@ -214,8 +214,6 @@ env:
   SWEBENCH_USE_MODAL: 'true'
   MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
   MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
-  # These b300 nodes are currently broken.
-  SALLOC_EXCLUDE: 'b300-005,b300-006,b300-017'
 
 permissions:
   contents: read

.github/workflows/speedbench-al.yml

@@ -75,7 +75,7 @@ permissions:
 env:
   HF_TOKEN: ${{ secrets.INFERENCEX_OFFICIAL_RO_HF_TOKEN }}
   HF_HUB_CACHE: '/mnt/hf_hub_cache/'
-  # Drive the single-node path in runners/launch_b300-nv.sh. MODEL is the HF id;
+  # Drive the single-node path in runners/launch_b300-dsxe.sh. MODEL is the HF id;
   # its basename (e.g. DeepSeek-V4-Pro) must be in the launcher's STAGED_MODELS so
   # the launcher resolves MODEL_PATH to the pre-staged local weights and mounts
   # them. The collector serves from MODEL_PATH (see SERVE_MODEL), so no download.

.github/workflows/test-process-result.yml

@@ -20,7 +20,7 @@ on:
       - 'runners/launch_gb300-nv.sh'
       - 'runners/launch_b200-nscale-slurm.sh'
       - 'runners/launch_b200-nscale-compat.sh'
-      - 'runners/launch_b300-nv.sh'
+      - 'runners/launch_b300-dsxe.sh'
       - 'runners/launch_h200-dgxc-slurm.sh'
       - 'runners/inject_srt_power_concurrencies.py'
       - 'utils/aggregate_power.py'

benchmarks/single_node/agentic/README.md

@@ -39,8 +39,8 @@ that runner label:
 
 ```yaml
 hardware:
-  cluster:b300-nv:
-    available-cpu-dram-mib: 2964436
+  cluster:b300-dsxe:
+    available-cpu-dram-mib: 3977095
     gpus-per-node: 8
 ```
 

benchmarks/single_node/agentic/dsv4_fp4_b300_vllm_mtp.sh

@@ -119,7 +119,7 @@ ROUTER_PID=""
 MOONCAKE_MASTER_PID=""
 
 # The generated TOTAL_CPU_DRAM_GB budget is proportional to allocated GPUs.
-# On cluster:b300-nv, dram-utilization=0.95 gives both DEP4 and DEP8 356 GB per
+# On cluster:b300-dsxe, dram-utilization=0.95 gives both DEP4 and DEP8 356 GB per
 # DP rank (1,424 GB and 2,849 GB total, respectively). TP arms remain GPU-resident.
 OFFLOAD_ARGS=()
 case "$KV_OFFLOAD_BACKEND" in

benchmarks/single_node/speedbench/dsr1_fp4_b300_vllm.sh

@@ -28,7 +28,7 @@
 #     shim (the default client-side template render already enables thinking).
 #
 # Checkpoint (B300 / Blackwell): NVFP4 build nvidia/DeepSeek-R1-0528-NVFP4-v2,
-# basename dsr1-fp4 on the runner (resolved by launch_b300-nv.sh).
+# basename dsr1-fp4 on the runner (resolved by launch_b300-dsxe.sh).
 #
 # Usage (inside the vLLM container, on a B300 node):
 #   export MODEL=/data/models/dsr1-fp4

configs/CONFIGS.md

@@ -152,13 +152,13 @@ hardware facts in the repository. It has two top-level sections:
 
 ```yaml
 labels:
-  cluster:b300-nv:
-    - b300-nv_01
-    - b300-nv_02
+  cluster:b300-dsxe:
+    - b300-dsxe_00
+    - b300-dsxe_01
 
 hardware:
-  cluster:b300-nv:
-    available-cpu-dram-mib: 2964436
+  cluster:b300-dsxe:
+    available-cpu-dram-mib: 3977095
     gpus-per-node: 8
 ```
 

configs/runners.yaml

@@ -99,6 +99,25 @@ labels:
     - gb200-nv_0
     - gb200-nv_1
     - gb200-nv_2
+  b300:
+    - b300-dsxe_00
+    - b300-dsxe_01
+    - b300-dsxe_02
+    - b300-dsxe_03
+    - b300-dsxe_04
+    - b300-dsxe_05
+    - b300-dsxe_06
+    - b300-dsxe_07
+    - b300-dsxe_08
+    - b300-dsxe_09
+    - b300-dsxe_10
+    - b300-dsxe_11
+    - b300-dsxe_12
+    - b300-dsxe_13
+    - b300-dsxe_14
+    - b300-dsxe_15
+    - b300-dsxe_16
+    - b300-dsxe_17
   gb300:
     - gb300-nv_0
     - gb300-nv_1
@@ -165,6 +184,25 @@ labels:
     - gb200-nv_1
     - gb200-nv_2
     - gb200-nv_3
+  cluster:b300-dsxe:
+    - b300-dsxe_00
+    - b300-dsxe_01
+    - b300-dsxe_02
+    - b300-dsxe_03
+    - b300-dsxe_04
+    - b300-dsxe_05
+    - b300-dsxe_06
+    - b300-dsxe_07
+    - b300-dsxe_08
+    - b300-dsxe_09
+    - b300-dsxe_10
+    - b300-dsxe_11
+    - b300-dsxe_12
+    - b300-dsxe_13
+    - b300-dsxe_14
+    - b300-dsxe_15
+    - b300-dsxe_16
+    - b300-dsxe_17
   cluster:gb300-nv:
     - gb300-nv_0
     - gb300-nv_1
@@ -221,6 +259,9 @@ hardware:
   cluster:b200-nscale:
     available-cpu-dram-mib: 2_063_920
     gpus-per-node: 8
+  cluster:b300-dsxe:
+    available-cpu-dram-mib: 3_977_095
+    gpus-per-node: 8
   cluster:gb200-nv:
     available-cpu-dram-mib: 860_160
     gpus-per-node: 4