[WIP] agentX v0.4

benchmarks/benchmark_lib.sh

@@ -924,8 +924,19 @@ resolve_trace_source() {
     # public-dataset loader names allowed by the inferencex-agentx-mvp
     # scenario. Used by recipes whose servers have non-default context
     # caps (e.g. minimaxm2.5 at max_model_len ~256k can't replay the
-    # unfiltered 052726 corpus and switches to the 256k-capped variant).
-    local loader="${WEKA_LOADER_OVERRIDE:-semianalysis_cc_traces_weka_with_subagents}"
+    # unfiltered corpus and switches to the 256k-capped variant), or
+    # by recipes that want to pin a specific corpus generation rather
+    # than ride the model-prefix-aware default below.
+    #
+    # Default (no override) is model-prefix-aware:
+    #   DSv4 recipes      -> 052726 (v5 corpus, the original baseline)
+    #   everything else   -> 060226 (v6 corpus, newer CC versions)
+    # DSv4 stays on 052726 for continuity with prior published baselines.
+    local default_loader="semianalysis_cc_traces_weka_with_subagents_060226"
+    if [[ "${MODEL_PREFIX:-}" == "dsv4" ]]; then
+        default_loader="semianalysis_cc_traces_weka_with_subagents"
+    fi
+    local loader="${WEKA_LOADER_OVERRIDE:-$default_loader}"
     local dataset
     case "$loader" in
         semianalysis_cc_traces_weka_with_subagents)
@@ -934,13 +945,19 @@ resolve_trace_source() {
         semianalysis_cc_traces_weka_with_subagents_256k)
             dataset="semianalysisai/cc-traces-weka-with-subagents-052726-256k"
             ;;
+        semianalysis_cc_traces_weka_with_subagents_060226)
+            dataset="semianalysisai/cc-traces-weka-with-subagents-060226"
+            ;;
+        semianalysis_cc_traces_weka_with_subagents_060226_256k)
+            dataset="semianalysisai/cc-traces-weka-with-subagents-060226-256k"
+            ;;
         *)
-            echo "Error: unknown WEKA_LOADER_OVERRIDE='$loader'. Allowed: semianalysis_cc_traces_weka_with_subagents, semianalysis_cc_traces_weka_with_subagents_256k" >&2
+            echo "Error: unknown WEKA_LOADER_OVERRIDE='$loader'. Allowed: semianalysis_cc_traces_weka_with_subagents, semianalysis_cc_traces_weka_with_subagents_256k, semianalysis_cc_traces_weka_with_subagents_060226, semianalysis_cc_traces_weka_with_subagents_060226_256k" >&2
             exit 1
             ;;
     esac
     TRACE_SOURCE_FLAG="--public-dataset $loader"
-    echo "Loading traces via aiperf public-dataset: $loader ($dataset)"
+    echo "Loading traces via aiperf public-dataset: $loader ($dataset) [MODEL_PREFIX=${MODEL_PREFIX:-unset}]"
     # Pre-download the dataset into the shared HF_HUB_CACHE (same mount used
     # for model weights) so subsequent runs read from cache instead of
     # re-downloading every job.

benchmarks/single_node/agentic/dsr1_fp4_b200.sh

@@ -17,7 +17,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -33,7 +43,7 @@ export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 
 python3 -m sglang.launch_server \
---model-path $MODEL \
+--model-path "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --trust-remote-code \

benchmarks/single_node/agentic/dsr1_fp4_mi355x.sh

@@ -16,7 +16,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 rocm-smi
 amd-smi || true
 
@@ -34,7 +44,7 @@ export ROCM_QUICK_REDUCE_QUANTIZATION=INT4
 export PYTHONNOUSERSITE=1
 
 python3 -m sglang.launch_server \
---model-path=$MODEL \
+--model-path=$MODEL_PATH --served-model-name=$MODEL \
 --host=0.0.0.0 \
 --port=$PORT \
 --trust-remote-code \

benchmarks/single_node/agentic/dsv4_fp4_b200_vllm.sh

@@ -38,7 +38,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -221,7 +231,7 @@ export VLLM_FLOAT32_MATMUL_PRECISION=high
 
 { set +x; } 2>/dev/null
 VLLM_CMD=(
-    vllm serve "$MODEL"
+    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
     --host 0.0.0.0
     --port "$PORT"
     --trust-remote-code

benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh

@@ -32,16 +32,35 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
+# Opt this recipe out of the DSv4 052726 default; use the v6 corpus.
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_with_subagents_060226
 resolve_trace_source
 install_agentic_deps
 
 # DeepSeek-V4-Pro weights are large; engine startup can exceed default 600s.
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 
+# vllm-project/vllm#43447: keep SWA prefix-cache tails sparsely so transient
+# sliding-window allocations don't evict useful prefix entries. 32k matches
+# the trace-replay tuning the PR author validated (0% -> 74% hit rate).
+# Requires the custom image (cquil/vllm-openai:*-7ead0a0f...) that carries
+# the patch; on stock images the env var is ignored.
+export VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768
+
 # ---- Server config ----------------------------------------------------------
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
@@ -113,7 +132,7 @@ export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 export VLLM_FLOAT32_MATMUL_PRECISION=high
 
-vllm serve "$MODEL" \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port "$PORT" \
 --trust-remote-code \

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh

@@ -31,7 +31,17 @@ if [ -n "${ROCR_VISIBLE_DEVICES:-}" ]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 rocm-smi || true
 amd-smi || true
 
@@ -135,7 +145,7 @@ fi
 
 echo "Starting sglang server..."
 python3 -m sglang.launch_server \
-    --model-path "$MODEL" \
+    --model-path "$MODEL_PATH" --served-model-name "$MODEL" \
     --host=0.0.0.0 \
     --port "$PORT" \
     "${PARALLEL_ARGS[@]}" \

benchmarks/single_node/agentic/dsv4_fp8_h200.sh

@@ -21,7 +21,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -40,7 +50,7 @@ export PYTHONNOUSERSITE=1
 
 # Per recipe: EP + DP=8 (no --tensor-parallel-size). TP from search space is
 # used for GPU allocation by the runner and as the DP size.
-vllm serve $MODEL \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --trust-remote-code \

benchmarks/single_node/agentic/glm5.1_fp4_mi355x.sh

@@ -19,7 +19,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 rocm-smi || true
 amd-smi || true
 
@@ -42,7 +52,7 @@ echo "Starting SGLang server..."
 export PYTHONNOUSERSITE=1
 
 python3 -m sglang.launch_server \
-    --model-path $MODEL \
+    --model-path "$MODEL_PATH" --served-model-name "$MODEL" \
     --host=0.0.0.0 \
     --port $PORT \
     --tensor-parallel-size $TP \

benchmarks/single_node/agentic/glm5_fp8_b200.sh

@@ -19,7 +19,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -39,7 +49,7 @@ export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 
 python3 -m sglang.launch_server \
---model-path=$MODEL \
+--model-path=$MODEL_PATH --served-model-name=$MODEL \
 --host=0.0.0.0 \
 --port=$PORT \
 --trust-remote-code \

benchmarks/single_node/agentic/gptoss_fp4_b200.sh

@@ -19,7 +19,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -53,7 +63,7 @@ export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 export VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8=1
 
-vllm serve $MODEL \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --config "$RESULT_DIR/config.yaml" \

benchmarks/single_node/agentic/gptoss_fp4_h100.sh

@@ -21,7 +21,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -57,7 +67,7 @@ export TORCH_CUDA_ARCH_LIST="9.0"
 export PYTHONNOUSERSITE=1
 export VLLM_MXFP4_USE_MARLIN=1
 
-vllm serve $MODEL \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --config "$RESULT_DIR/config.yaml" \

benchmarks/single_node/agentic/gptoss_fp4_h200.sh

@@ -21,7 +21,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 nvidia-smi
 
 # ---- Resolve traces and install deps ----------------------------------------
@@ -57,7 +67,7 @@ export TORCH_CUDA_ARCH_LIST="9.0"
 export PYTHONNOUSERSITE=1
 export VLLM_MXFP4_USE_MARLIN=1
 
-vllm serve $MODEL \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --config "$RESULT_DIR/config.yaml" \

benchmarks/single_node/agentic/gptoss_fp4_mi300x.sh

@@ -21,7 +21,17 @@ if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
 fi
 
-if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+# `hf download` creates the target dir if missing and is itself idempotent.
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# Either way, MODEL_PATH is what the server is launched with.
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
 rocm-smi
 amd-smi || true
 
@@ -65,7 +75,7 @@ esac
 
 echo "Starting vllm server..."
 
-vllm serve $MODEL \
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
 --host 0.0.0.0 \
 --port $PORT \
 --attention-backend ROCM_AITER_UNIFIED_ATTN \