Use local MODEL_PATH for B200 DGXC single-node bench scripts (#1317)

* Use local MODEL_PATH for B200 DGXC single-node bench scripts

Hoist the MODEL_PATH override out of the multinode-only block so single-node
launches on the B200 DGXC cluster also load models from /lustre/fsw/models
instead of pulling through the HF hub cache. Single-node now mounts
MODEL_PATH into the container and exports MODEL=$MODEL_PATH so existing
bench scripts pick up the local directory via --model-path.

Guard `hf download "$MODEL"` in 89 single-node bench scripts using the
leading-slash check that already exists in agentic/*.sh and
dsv4_fp4_b300_sglang*.sh, so the download is skipped when MODEL is a local
path. Print the /lustre/fsw/models listing in the error path when an
unknown MODEL_PREFIX/PRECISION combo is requested. Preserve the
multinode dsv4-only-with-dynamo-vllm constraint as an explicit guard
since hoisting dropped the framework filter from the path-resolution
branch.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Remove stale b200-dgxc_10..16 runners from b200 pool

The B200 DGXC cluster only has nodes 00-09; entries 10-16 in the b200
runner pool were stale and would cause jobs to queue against runners
that don't exist.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Update b200-multinode runner pool to slurm_7..9

The actual b200 multinode pool is b200-dgxc-slurm_7, _8, _9 — _6 was
stale.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Switch B200 DGXC MODEL_PATH resolver to /raid/models + cover all configs

Lustre-over-TCP at /lustre/fsw/models is ~98% full and slower than the
per-node local RAID array. Point MODEL_PATH at /raid/models/* instead.

Also add resolver branches for every (model-prefix, precision) combo that
nvidia-master.yaml declares as single-node runner: b200 — qwen3.5
(bf16/fp8/fp4), glm5 (fp8/fp4), kimik2.5 (int4/fp4), minimaxm2.5
(fp8/fp4), and gptoss (fp4). Previously these 16 configs would hard-fail
at `exit 1` when they landed on a b200-dgxc runner because the override
table only covered dsr1 and dsv4.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Revert /raid model paths back to /lustre/fsw/models

/raid/models/* is populated on only a subset of dgxc compute nodes today
(survey via srun showed 6 of 10 gpu-2 nodes had the dsr1 dir). The
preview verification run hit `enroot-mount: failed to mount
/raid/models/dsr1-0528-nvfp4-v2 ... No such file or directory` on
gpu-2-2.

Point all 13 resolver branches back at /lustre/fsw/models/* until /raid
staging is reliable across the fleet.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: cquil11

.github/configs/runners.yaml

@@ -79,17 +79,10 @@ b200:
 - 'b200-dgxc_07'
 - 'b200-dgxc_08'
 - 'b200-dgxc_09'
-- 'b200-dgxc_10'
-- 'b200-dgxc_11'
-- 'b200-dgxc_12'
-- 'b200-dgxc_13'
-- 'b200-dgxc_14'
-- 'b200-dgxc_15'
-- 'b200-dgxc_16'
 b200-multinode:
-- 'b200-dgxc-slurm_6'
 - 'b200-dgxc-slurm_7'
 - 'b200-dgxc-slurm_8'
+- 'b200-dgxc-slurm_9'
 mi300x:
 - 'mi300x-amds_00'
 - 'mi300x-amds_01'

benchmarks/single_node/dsr1_fp4_b200.sh

@@ -16,7 +16,7 @@ if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 fi
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 nvidia-smi
 

benchmarks/single_node/dsr1_fp4_b200_trt.sh

@@ -20,7 +20,7 @@ fi
 
 echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 # ========= Determine other parameters based on ISL, OSL, CONC =========
 CUDA_GRAPH_MAX_BATCH_SIZE=$CONC

benchmarks/single_node/dsr1_fp4_b200_trt_mtp.sh

@@ -20,7 +20,7 @@ fi
 
 echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 # ========= Determine MOE_BACKEND and MTP based on DP_ATTENTION =========
 MOE_BACKEND="TRTLLM"

benchmarks/single_node/dsr1_fp4_b300.sh

@@ -20,7 +20,7 @@ if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 fi
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 nvidia-smi
 

benchmarks/single_node/dsr1_fp4_mi355x.sh

@@ -15,7 +15,7 @@ if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 fi
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 export SGLANG_USE_AITER=1
 export ROCM_QUICK_REDUCE_QUANTIZATION=INT4

benchmarks/single_node/dsr1_fp8_b200.sh

@@ -18,7 +18,7 @@ fi
 
 nvidia-smi
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 export SGL_ENABLE_JIT_DEEPGEMM=false
 export SGLANG_ENABLE_FLASHINFER_GEMM=true

benchmarks/single_node/dsr1_fp8_b200_mtp.sh

@@ -18,7 +18,7 @@ fi
 
 nvidia-smi
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 export SGLANG_ENABLE_JIT_DEEPGEMM=false
 

benchmarks/single_node/dsr1_fp8_b200_trt.sh

@@ -20,7 +20,7 @@ fi
 
 echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 # temporary, avoids risk of OOM error
 export TLLM_OVERRIDE_LAYER_NUM=61

benchmarks/single_node/dsr1_fp8_b200_trt_mtp.sh

@@ -20,7 +20,7 @@ fi
 
 echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
 
-hf download "$MODEL"
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
 
 # ========= Determine other parameters based on ISL, OSL, CONC =========
 MOE_BACKEND="TRTLLM"