refactor: split build_gpu_mem_args into engine-specific functions (#7916)

Signed-off-by: Keiven Chang <keivenchang@users.noreply.github.com>

examples/backends/sglang/launch/agg.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Default values
@@ -55,7 +55,7 @@ if [ "$ENABLE_OTEL" = true ]; then
     TRACE_ARGS+=(--enable-trace --otlp-traces-endpoint localhost:4317)
 fi
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated Serving" "$MODEL" "$HTTP_PORT"

examples/backends/sglang/launch/agg_embed.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Default values
@@ -40,7 +40,7 @@ while [[ $# -gt 0 ]]; do
     esac
 done
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner --no-curl "Launching Embedding Worker" "$MODEL" "$HTTP_PORT"

examples/backends/sglang/launch/agg_router.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Parse command line arguments
@@ -54,7 +54,7 @@ fi
 
 MODEL="Qwen/Qwen3-0.6B"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated + KV Routing (2 GPUs)" "$MODEL" "$HTTP_PORT"

examples/backends/sglang/launch/disagg.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Parse command line arguments
@@ -48,7 +48,7 @@ fi
 
 MODEL="Qwen/Qwen3-0.6B"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 DISAGG_BOOTSTRAP_PORT="${DYN_DISAGG_BOOTSTRAP_PORT:-12345}"
 

examples/backends/sglang/launch/disagg_same_gpu.sh

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 # Disaggregated prefill/decode on a SINGLE GPU.
-# Per-worker VRAM is controlled via build_gpu_mem_args (see gpu_utils.sh).
+# Per-worker VRAM is controlled via build_sglang_gpu_mem_args (see gpu_utils.sh).
 # Override individual knobs (CONTEXT_LENGTH, MAX_RUNNING_REQUESTS) via env vars.
 #
 # Measured reference (Qwen/Qwen3-0.6B, --context-length 4096, RTX 6000 Ada 48 GiB):
@@ -25,7 +25,7 @@ MODEL="Qwen/Qwen3-0.6B"
 CONTEXT_LENGTH="${CONTEXT_LENGTH:-4096}"
 MAX_RUNNING_REQUESTS="${MAX_RUNNING_REQUESTS:-2}"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang --workers-per-gpu 2)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"
 

examples/backends/sglang/launch/multimodal_disagg.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Default values
@@ -99,7 +99,7 @@ DYN_ENCODE_GPU_MEM=${DYN_ENCODE_GPU_MEM:-0.9}
 DYN_PREFILL_GPU_MEM=${DYN_PREFILL_GPU_MEM:-0.9}
 DYN_DECODE_GPU_MEM=${DYN_DECODE_GPU_MEM:-0.9}
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang --workers-per-gpu 3)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 ENCODE_EXTRA_ARGS=""
 PREFILL_EXTRA_ARGS=""

examples/backends/sglang/launch/multimodal_epd.sh

@@ -9,7 +9,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_gpu_mem_args
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_sglang_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Default values
@@ -94,7 +94,7 @@ fi
 DYN_ENCODE_GPU_MEM=${DYN_ENCODE_GPU_MEM:-0.9}
 DYN_WORKER_GPU_MEM=${DYN_WORKER_GPU_MEM:-0.9}
 
-GPU_MEM_ARGS=$(build_gpu_mem_args sglang --workers-per-gpu 2)
+GPU_MEM_ARGS=$(build_sglang_gpu_mem_args)
 
 ENCODE_EXTRA_ARGS=""
 WORKER_EXTRA_ARGS=""

examples/backends/trtllm/launch/disagg_same_gpu.sh

@@ -4,13 +4,13 @@
 #
 # Disaggregated prefill/decode on a SINGLE GPU.
 # Per-worker VRAM is controlled via env vars (MAX_SEQ_LEN, MAX_CONCURRENT_SEQS).
-# TODO: unify with build_gpu_mem_args once trtllm --override-engine-args JSON
+# TODO: unify with build_trtllm_override_args_with_mem once trtllm --override-engine-args JSON
 # merging is supported.
 #
 # NOTE — trtllm fraction semantics differ from vllm/sglang:
 #   vllm/sglang:  fraction of TOTAL VRAM  (weights + KV + activations all inside)
 #   trtllm:       fraction of FREE  VRAM  (KV cache only, after model load)
-# build_gpu_mem_args handles this — see gpu_utils.sh / gpu_utils.md.
+# build_vllm_gpu_mem_args / build_sglang_gpu_mem_args handle this — see gpu_utils.sh / gpu_utils.md.
 #
 # Measured reference (Qwen/Qwen3-0.6B, --max-seq-len 4096, RTX 6000 Ada 48 GiB):
 #   estimate (from gpu_utils.sh) : ~8.0 GiB per worker (~16.0 GiB total)
@@ -30,7 +30,7 @@ MODEL="Qwen/Qwen3-0.6B"
 MAX_SEQ_LEN="${MAX_SEQ_LEN:-4096}"
 MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 
-# TODO: unify with build_gpu_mem_args once trtllm --override-engine-args JSON
+# TODO: unify with build_trtllm_override_args_with_mem once trtllm --override-engine-args JSON
 # merging is supported.
 GPU_MEM_FRACTION="${GPU_MEM_FRACTION:-}"
 

examples/backends/vllm/launch/agg.sh

@@ -8,7 +8,7 @@ set -e
 trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # gpu_gb_to_total_fraction
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"   # build_vllm_gpu_mem_args
 source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait_any_exit
 
 # Default model
@@ -33,7 +33,7 @@ done
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
 MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated Serving (1 GPU)" "$MODEL" "$HTTP_PORT"

examples/backends/vllm/launch/agg_kvbm_router.sh

@@ -24,7 +24,7 @@ python -m dynamo.frontend \
 # run workers with KVBM enabled
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
 # Each worker needs unique ZMQ ports to avoid KVBM coordination conflicts
-# TODO: use build_gpu_mem_args to measure VRAM instead of hardcoded fractions
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of hardcoded fractions
 DYN_KVBM_LEADER_ZMQ_PUB_PORT=56001 \
 DYN_KVBM_LEADER_ZMQ_ACK_PORT=56002 \
 CUDA_VISIBLE_DEVICES=0 DYN_KVBM_CPU_CACHE_GB=2 \

examples/backends/vllm/launch/agg_lmcache.sh

@@ -17,7 +17,7 @@ MODEL="Qwen/Qwen3-0.6B"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
 MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated Serving + LMCache (1 GPU)" "$MODEL" "$HTTP_PORT"

examples/backends/vllm/launch/agg_lmcache_multiproc.sh

@@ -27,7 +27,7 @@ MODEL="Qwen/Qwen3-0.6B"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
 MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated + LMCache + Multiproc (1 GPU)" "$MODEL" "$HTTP_PORT"

examples/backends/vllm/launch/agg_multimodal.sh

@@ -97,7 +97,7 @@ case "$MODEL_NAME" in
         MODEL_EXTRA_ARGS="--tensor-parallel-size=8" ;;
 esac
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 # Start vLLM worker with vision model
 # --enforce-eager: Quick deployment (remove for production)

examples/backends/vllm/launch/agg_request_planes.sh

@@ -50,7 +50,7 @@ MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 export DYN_REQUEST_PLANE=$REQUEST_PLANE
 echo "Using request plane mode: $REQUEST_PLANE"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
 print_launch_banner "Launching Aggregated Serving + Request Planes (1 GPU)" "$MODEL" "$HTTP_PORT"

examples/backends/vllm/launch/agg_router.sh

@@ -29,7 +29,7 @@ python -m dynamo.frontend \
 #
 # If multiple workers are launched, they must not share the same system/metrics port.
 # Use DYN_SYSTEM_PORT{1,2} so tests/launchers can provide a simple numbered port set.
-# TODO: use build_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
 #
 DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \

examples/backends/vllm/launch/agg_router_approx.sh

@@ -23,7 +23,7 @@ python -m dynamo.frontend \
 #
 # If multiple workers are launched, they must not share the same system/metrics port.
 # Use DYN_SYSTEM_PORT{1,2} so tests/launchers can provide a simple numbered port set.
-# TODO: use build_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
 
 DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \

examples/backends/vllm/launch/agg_spec_decoding.sh

@@ -21,7 +21,7 @@ python -m dynamo.frontend --http-port="$HTTP_PORT" &
 # 2. Speculative Main Worker
 # ---------------------------
 # This runs the main model with EAGLE as the draft model for speculative decoding
-# TODO: use build_gpu_mem_args to measure VRAM instead of hardcoded fractions
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of hardcoded fractions
 DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=8081 \
 CUDA_VISIBLE_DEVICES=0 python -m dynamo.vllm \
     --model "$MODEL" \

examples/backends/vllm/launch/disagg.sh

@@ -18,7 +18,7 @@ print_launch_banner "Launching Disaggregated Serving (2 GPUs)" "$MODEL" "$HTTP_P
 python -m dynamo.frontend &
 
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
-# TODO: use build_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of relying on vLLM defaults
 DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT1:-8081} \
 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \
     --model "$MODEL" \

examples/backends/vllm/launch/disagg_multimodal_e_pd.sh

@@ -79,7 +79,7 @@ python -m dynamo.frontend &
 EXTRA_ARGS=""
 
 # GPU assignments (override via environment variables)
-# TODO: use build_gpu_mem_args to measure VRAM instead of hardcoded fractions
+# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of hardcoded fractions
 # In single-GPU mode both workers share the same GPU.
 if [[ "$SINGLE_GPU" == "true" ]]; then
     DYN_ENCODE_WORKER_GPU=${DYN_ENCODE_WORKER_GPU:-0}

examples/backends/vllm/launch/disagg_same_gpu.sh

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 # Disaggregated prefill/decode on a SINGLE GPU.
-# Per-worker VRAM is controlled via build_gpu_mem_args (see gpu_utils.sh).
+# Per-worker VRAM is controlled via build_vllm_gpu_mem_args (see gpu_utils.sh).
 # Override individual knobs (MAX_MODEL_LEN, MAX_CONCURRENT_SEQS) via env vars.
 #
 # Measured reference (Qwen/Qwen3-0.6B, --max-model-len 4096, RTX 6000 Ada 48 GiB):
@@ -25,7 +25,7 @@ MODEL="Qwen/Qwen3-0.6B"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
 MAX_CONCURRENT_SEQS="${MAX_CONCURRENT_SEQS:-2}"
 
-GPU_MEM_ARGS=$(build_gpu_mem_args vllm --workers-per-gpu 2)
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"