test(serve): add aggregated_spec_decoding launch scenario

examples/backends/vllm/launch/agg_spec_decoding.sh

@@ -6,6 +6,12 @@ trap 'echo Cleaning up...; kill 0' EXIT
 
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
 source "$SCRIPT_DIR/../../../common/launch_utils.sh"
+source "$SCRIPT_DIR/../../../common/gpu_utils.sh"  # build_vllm_gpu_mem_args
+
+# When the serve test sets requested_vllm_kv_cache_bytes, this emits a
+# --kv-cache-memory-bytes cap so VRAM is bounded (GPU-size-independent) and the
+# test can run in the VRAM-aware parallel stage. Empty for manual runs.
+GPU_MEM_ARGS=$(build_vllm_gpu_mem_args)
 
 MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"
 HTTP_PORT="${DYN_HTTP_PORT:-8000}"
@@ -21,8 +27,7 @@ python -m dynamo.frontend --http-port="$HTTP_PORT" &
 # 2. Speculative Main Worker
 # ---------------------------
 # This runs the main model with EAGLE as the draft model for speculative decoding
-# TODO: use build_vllm_gpu_mem_args to measure VRAM instead of hardcoded fractions
-DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=8081 \
+DYN_SYSTEM_ENABLED=true DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT:-8081} \
 CUDA_VISIBLE_DEVICES=0 python -m dynamo.vllm \
     --model "$MODEL" \
     --enforce-eager \
@@ -32,7 +37,7 @@ CUDA_VISIBLE_DEVICES=0 python -m dynamo.vllm \
         "num_speculative_tokens": 2,
         "method": "eagle3"
     }' \
-    --gpu-memory-utilization 0.8 &
+    --gpu-memory-utilization 0.8 $GPU_MEM_ARGS "$@" &
 
 # Exit on first worker failure; kill 0 in the EXIT trap tears down the rest
 wait_any_exit

tests/serve/test_vllm.py

@@ -160,6 +160,42 @@ class VLLMConfig(EngineConfig):
             metric_payload_default(min_num_requests=6, backend="vllm"),
         ],
     ),
+    # Speculative decoding: Llama-3.1-8B main model with an EAGLE3 draft model
+    # (see launch/agg_spec_decoding.sh). The base model is gated on HF, so this
+    # needs HF_TOKEN set and only runs where the token + VRAM are available.
+    # Nightly-only: the 8B base plus EAGLE3 draft model is intentionally outside pre-merge CI.
+    "aggregated_spec_decoding": VLLMConfig(
+        name="aggregated_spec_decoding",
+        directory=vllm_dir,
+        script_name="agg_spec_decoding.sh",
+        marks=[
+            pytest.mark.gpu_1,
+            # Also predownload the EAGLE3 draft: CI workers run HF_HUB_OFFLINE=True
+            # and only the base cfg.model is auto-registered, so the draft repo
+            # can't be resolved offline without this.
+            pytest.mark.model("yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"),
+            # Profiled with a 1 GiB KV cap (peak ~18.6 GiB: 8B weights + EAGLE3
+            # draft + capped KV). The byte cap (build_vllm_gpu_mem_args) makes
+            # this GPU-size-independent, so it fits the 24 GiB parallel stage.
+            pytest.mark.profiled_vram_gib(20.0),
+            pytest.mark.requested_vllm_kv_cache_bytes(1_073_741_824),
+            pytest.mark.timeout(900),
+            pytest.mark.nightly,
+        ],
+        model="meta-llama/Meta-Llama-3.1-8B-Instruct",
+        # 8B weights leave little KV room on the 24 GiB gpu_1 lane; cap context (payloads are tiny).
+        script_args=["--max-model-len", "4096"],
+        request_payloads=[
+            chat_payload_default(),
+            chat_payload(
+                "What is the capital of France? Answer in one word.",
+                repeat_count=1,
+                expected_response=["Paris"],
+                temperature=0.0,
+                max_tokens=16,
+            ),
+        ],
+    ),
     "aggregated_unified": VLLMConfig(
         name="aggregated_unified",
         directory=vllm_dir,