fix(sglang/dsv4/8k1k recipes): set cpus-per-task=144 and mem=0

slurm assigns 1 CPU/task by default; `scontrol show job 613` from a
running CI job confirmed `NumCPUs=4 NumTasks=4 CPUs/Task=1` with 4
nodes — one core per worker. The dynamo `hash:` cold source install
rebuilds ~500 rust crates (kube-client, tonic, hf-hub, image codecs
ravif/exr, the pyo3 stack) and at one core takes 30+ min just for the
cold build, which dominates total CI time even with the new
`/configs/dynamo-wheels` cache (the cache only helps after the first
cold run).

Match yangminl's working manual setup on the same gb300-cw cluster
(`/mnt/home/yangminl/srt-slurm/recipes/dsv4-pro/sglang/gb300-fp4/all-dynamo.yaml`)
which sets:
  sbatch_directives:
    cpus-per-task: "144"
    mem: "0"

cargo then gets the full 144-core GB300 host and finishes maturin in a
few minutes; mem=0 hands the worker the entire node's RAM so the
dynamo build + DSV4-Pro 671B FP4 weight load fit without OOM.

Author: ch-wan

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-2p1d-dep4-dep8.yaml

@@ -34,11 +34,17 @@ dynamo:
 slurm:
   time_limit: "8:00:00"
 
-# Without cpus-per-task slurm gives 1 CPU/task; the dynamo cold source
-# build (~500 rust crates including ravif/exr/zip) is otherwise serial
-# and takes 30+ min. Match yangminl's all-dynamo.yaml which uses 144.
+# Match yangminl's working all-dynamo.yaml on the same gb300-cw cluster:
+#   cpus-per-task=144  — without this slurm hands out 1 CPU/task, which
+#     turns the dynamo `hash:` cold source build (~500 rust crates,
+#     ravif/exr/zip/pyo3 stack) into a 30+ min serial compile. With 144
+#     cargo finishes in ~5 min.
+#   mem=0              — slurm's "give the whole node's memory"; needed
+#     for sglang loading 671B FP4 weights + dynamo build at the same
+#     time without OOM.
 sbatch_directives:
   cpus-per-task: "144"
+  mem: "0"
 
 health_check:
   max_attempts: 1440

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-7p1d-dep4-dep8.yaml

@@ -34,11 +34,17 @@ dynamo:
 slurm:
   time_limit: "8:00:00"
 
-# Without cpus-per-task slurm gives 1 CPU/task; the dynamo cold source
-# build (~500 rust crates including ravif/exr/zip) is otherwise serial
-# and takes 30+ min. Match yangminl's all-dynamo.yaml which uses 144.
+# Match yangminl's working all-dynamo.yaml on the same gb300-cw cluster:
+#   cpus-per-task=144  — without this slurm hands out 1 CPU/task, which
+#     turns the dynamo `hash:` cold source build (~500 rust crates,
+#     ravif/exr/zip/pyo3 stack) into a 30+ min serial compile. With 144
+#     cargo finishes in ~5 min.
+#   mem=0              — slurm's "give the whole node's memory"; needed
+#     for sglang loading 671B FP4 weights + dynamo build at the same
+#     time without OOM.
 sbatch_directives:
   cpus-per-task: "144"
+  mem: "0"
 
 health_check:
   max_attempts: 1440