dsv4 sglang agentic: enable out-of-window SWA slot release and bump image to nightly-20260707

benchmarks/single_node/agentic/dsv4_fp4_b200_sglang.sh

@@ -56,6 +56,9 @@ install_agentic_deps
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
+export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
+export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+
 CACHE_ARGS=()
 if require_agentic_kv_offload_backend hicache; then
     # DeepSeek V4 HiCache currently rejects --hicache-size and supports
@@ -72,7 +75,6 @@ if require_agentic_kv_offload_backend hicache; then
     HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_through}"
     HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
     HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-page_first_direct}"
-    export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
     CACHE_ARGS=(
         --enable-hierarchical-cache
         --hicache-ratio "$HICACHE_RATIO"

benchmarks/single_node/agentic/dsv4_fp4_b300_sglang.sh

@@ -56,6 +56,9 @@ install_agentic_deps
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
+export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
+export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
+
 CACHE_ARGS=()
 if require_agentic_kv_offload_backend hicache; then
     # DeepSeek V4 HiCache currently rejects --hicache-size and supports
@@ -77,7 +80,6 @@ if require_agentic_kv_offload_backend hicache; then
     HICACHE_WRITE_POLICY="${HICACHE_WRITE_POLICY:-write_back}"
     HICACHE_IO_BACKEND="${HICACHE_IO_BACKEND:-direct}"
     HICACHE_MEM_LAYOUT="${HICACHE_MEM_LAYOUT:-page_first_direct}"
-    export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
     CACHE_ARGS=(
         --enable-hierarchical-cache
         --hicache-ratio "$HICACHE_RATIO"

configs/nvidia-master.yaml

@@ -13767,7 +13767,7 @@ minimaxm3-fp8-h200-vllm-agentic:
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: mooncake, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
 
 dsv4-fp4-b200-sglang-agentic-hicache:
-  image: lmsysorg/sglang:v0.5.13-cu130
+  image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b200-dgxc
@@ -13789,7 +13789,7 @@ dsv4-fp4-b200-sglang-agentic-hicache:
 # covers the middle/high-interactivity range omitted by the one-decode DEP
 # throughput curves below. Each engine start carries at most four concurrencies.
 dsv4-fp4-b300-sglang-agentic-hicache:
-  image: lmsysorg/sglang:v0.5.13-cu130
+  image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b300-nv

perf-changelog.yaml

@@ -4643,3 +4643,10 @@
     - "Recipes sourced from NVIDIA/srt-slurm branch sa-submission-q2-2026 (gb300_nvfp4 MTP recipes)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1799
 
+- config-keys:
+    - dsv4-fp4-b200-sglang-agentic-hicache
+    - dsv4-fp4-b300-sglang-agentic-hicache
+  description:
+    - "Update SGLang image from lmsysorg/sglang:v0.5.13-cu130 to lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+    - "Enable SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 unconditionally (previously hicache-only) and SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 to free out-of-window SWA KV slots during chunked prefill, relieving SWA pool pressure and restoring prefix-cache hit rate on multi-turn agentic workloads"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2112

runners/launch_b200-dgxc.sh

@@ -430,24 +430,6 @@ else
     salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
-    # DSv4 is also staged on the compute nodes' local RAID. Loading the 806 GB
-    # checkpoint independently from Lustre on every TP rank leaves the loader
-    # threads blocked in Lustre I/O for hours. Select the local copy only after
-    # Slurm assigns a node, and retain the shared-Lustre path as a fallback for
-    # nodes whose local staging is incomplete.
-    if [[ "$MODEL_PREFIX" == "dsv4" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "sglang" ]]; then
-        LOCAL_MODEL_PATH=/raid/models/DeepSeek-V4-Pro-NVFP4
-        if srun --jobid="$JOB_ID" bash -c \
-            'test -f "$1/config.json" && test -f "$1/model.safetensors.index.json" && test "$(find "$1" -maxdepth 1 -name "model-*.safetensors" | wc -l)" -eq 64' \
-            _ "$LOCAL_MODEL_PATH"; then
-            export MODEL_PATH="$LOCAL_MODEL_PATH"
-            export MODEL="$MODEL_PATH"
-            echo "Using node-local DSv4 checkpoint: $MODEL_PATH"
-        else
-            echo "Node-local DSv4 checkpoint unavailable; using shared checkpoint: $MODEL_PATH"
-        fi
-    fi
-
     # Use flock to serialize concurrent imports to the same squash file
     # Override ENROOT_CACHE_PATH to avoid permission issues with system-wide cache on worker nodes
     srun --jobid=$JOB_ID bash -c "