feat(vla): integrate LingBot VLA v2 structured serving and benchmarks

.gitignore

@@ -158,6 +158,7 @@ INSTALL_HYS.md
 AGENTS.md
 _version.py.mcp.json
 telefuser/_version.py
+!telefuser/pipelines/lingbot_vla_v2/assets/*.json
 # LingBot regression example assets
 !examples/data/lingbot_world_fast/image.jpg
 !examples/data/lingbot_world_fast/poses.npy

benchmarks/telefuser_aiperf/README.md

@@ -92,6 +92,31 @@ Available batch configs:
 | `configs/video_generation_rate.yaml` | Poisson-arrival load |
 | `configs/video_generation_wan21_i2v_480p_compare.yaml` | Fixed Wan2.1 I2V comparison |
 
+## LingBot-VLA v2 Structured Actions
+
+Start the native VLA service from its isolated model environment, then run the AIPerf workload from the repository
+root:
+
+```bash
+bash benchmarks/telefuser_aiperf/scripts/run_vla_structured_bench.sh
+```
+
+The repository-owned `telefuser_vla_structured` endpoint and `telefuser_structured_http` transport submit
+`POST /v1/tasks/structured`, poll `GET /v1/tasks/{task_id}/status`, and pass request latency, throughput, success,
+trace, and server metric facts into AIPerf's normal warmup and aggregation pipeline. Defaults are two excluded warmup
+requests followed by 20 measured requests at concurrency one. Override them without changing the checked-in config:
+
+```bash
+TELEFUSER_AIPERF_REQUESTS=100 \
+TELEFUSER_AIPERF_CONCURRENCY=2 \
+  bash benchmarks/telefuser_aiperf/scripts/run_vla_structured_bench.sh
+```
+
+Each terminal result is required to contain a finite `50x55` action chunk and the frozen structured result fields.
+The adapter retains an action hash, bounds, dimensions, verification status, target inference time, and peak memory;
+it does not copy full action arrays or Base64 cameras into AIPerf response records. This validates service execution
+and normalized action structure, not physical robot control semantics.
+
 ## LingBot-World v2 Streaming
 
 The v2 pipeline expects the following files below `TF_MODEL_ZOO_PATH`:
@@ -264,10 +289,12 @@ AIPerf environment first, then run the checks from the repository root:
 PYTHONPATH=benchmarks/telefuser_aiperf \
   .venv-aiperf/bin/python -m pytest \
   benchmarks/telefuser_aiperf/tests/test_livekit_adapter.py \
-  benchmarks/telefuser_aiperf/tests/test_sglang_adapter.py
+  benchmarks/telefuser_aiperf/tests/test_sglang_adapter.py \
+  benchmarks/telefuser_aiperf/tests/test_vla_structured.py
 
 bash -n \
   scripts/setup_aiperf.sh \
   benchmarks/telefuser_aiperf/scripts/run_stream_bench.sh \
-  benchmarks/telefuser_aiperf/scripts/run_sglang_lingbot_world_v2_4gpu.sh
+  benchmarks/telefuser_aiperf/scripts/run_sglang_lingbot_world_v2_4gpu.sh \
+  benchmarks/telefuser_aiperf/scripts/run_vla_structured_bench.sh
 ```

benchmarks/telefuser_aiperf/configs/vla_structured_e2e.yaml

@@ -0,0 +1,49 @@
+# yaml-language-server: $schema=../../aiperf/src/aiperf/config/schema/aiperf-config.schema.json
+
+schemaVersion: "2.0"
+
+randomSeed: 42
+
+benchmark:
+  model: lingbot-vla-v2-6b
+
+  endpoint:
+    url: ${TELEFUSER_AIPERF_URL:http://127.0.0.1:18080}
+    type: telefuser_vla_structured
+    transport: telefuser_structured_http
+    timeout: ${TELEFUSER_AIPERF_TIMEOUT:120}
+
+  tokenizer:
+    name: builtin
+
+  dataset:
+    type: file
+    path: ./benchmarks/telefuser_aiperf/data/vla_structured.jsonl
+    format: single_turn
+    sampling: sequential
+
+  warmup:
+    type: concurrency
+    concurrency: 1
+    requests: ${TELEFUSER_AIPERF_WARMUP_REQUESTS:2}
+    excludeFromResults: true
+
+  profiling:
+    type: concurrency
+    concurrency: ${TELEFUSER_AIPERF_CONCURRENCY:1}
+    requests: ${TELEFUSER_AIPERF_REQUESTS:20}
+    duration: ${TELEFUSER_AIPERF_DURATION:3600}
+    gracePeriod: ${TELEFUSER_AIPERF_GRACE_PERIOD:120}
+
+  artifacts:
+    dir: ./artifacts/telefuser_aiperf/vla_structured
+    summary: [json]
+    records: [jsonl]
+    showTraceTiming: true
+    trace: true
+
+  serverMetrics:
+    enabled: ${TELEFUSER_AIPERF_SERVER_METRICS:true}
+    urls:
+      - ${TELEFUSER_AIPERF_METRICS_URL:http://127.0.0.1:18080/v1/service/metrics}
+    formats: [json, csv]

benchmarks/telefuser_aiperf/data/vla_structured.jsonl

@@ -0,0 +1 @@
+{"text":"pick up the object","image":"examples/data/101235-video-720_0.png","extra":{"state":[0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0],"seed":7}}

benchmarks/telefuser_aiperf/scripts/run_vla_structured_bench.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"
+cd "${ROOT_DIR}"
+
+CONFIG_PATH="${1:-benchmarks/telefuser_aiperf/configs/vla_structured_e2e.yaml}"
+SERVER_URL="${TELEFUSER_AIPERF_URL:-http://127.0.0.1:18080}"
+HEALTH_URL="${TELEFUSER_AIPERF_HEALTH_URL:-${SERVER_URL}/v1/service/ready}"
+DEFAULT_PYTHON="${ROOT_DIR}/.venv-aiperf/bin/python"
+ADAPTER_ROOT="${ROOT_DIR}/benchmarks/telefuser_aiperf"
+AIPERF_PYTHON="${TELEFUSER_AIPERF_PYTHON:-${DEFAULT_PYTHON}}"
+
+if [[ ! -x "${AIPERF_PYTHON}" ]]; then
+    echo "The isolated AIPerf environment is unavailable. Run: bash scripts/setup_aiperf.sh" >&2
+    exit 1
+fi
+
+if command -v curl >/dev/null 2>&1; then
+    echo "Checking TeleFuser VLA readiness: ${HEALTH_URL}"
+    curl --noproxy '*' --fail --silent --show-error "${HEALTH_URL}" >/dev/null
+fi
+
+export PYTHONPATH="${ADAPTER_ROOT}${PYTHONPATH:+:${PYTHONPATH}}"
+exec "${AIPERF_PYTHON}" -m telefuser_aiperf.cli profile --config "${CONFIG_PATH}"

benchmarks/telefuser_aiperf/telefuser_aiperf/__init__.py

@@ -6,6 +6,12 @@
 
 from telefuser_aiperf.adapter import TeleFuserLiveKitAdapter
 from telefuser_aiperf.sglang_adapter import SGLangRealtimeAdapter
+from telefuser_aiperf.vla_structured import (
+    ENDPOINT_METADATA,
+    TRANSPORT_METADATA,
+    TeleFuserStructuredHttpTransport,
+    TeleFuserVlaStructuredEndpoint,
+)
 
 
 def register_adapters(*, replace: bool = False) -> None:
@@ -23,4 +29,43 @@ def register_adapters(*, replace: bool = False) -> None:
     )
 
 
-__all__ = ["SGLangRealtimeAdapter", "TeleFuserLiveKitAdapter", "register_adapters"]
+def register_plugins(*, replace: bool = False) -> None:
+    """Register repository-owned AIPerf batch endpoint and transport plugins."""
+    from aiperf.plugin import plugins
+    from aiperf.plugin.enums import EndpointType, TransportType
+
+    if "telefuser_vla_structured" not in EndpointType:
+        EndpointType.register("TELEFUSER_VLA_STRUCTURED", "telefuser_vla_structured")
+    if "telefuser_structured_http" not in TransportType:
+        TransportType.register("TELEFUSER_STRUCTURED_HTTP", "telefuser_structured_http")
+
+    definitions = (
+        (
+            "endpoint",
+            "telefuser_vla_structured",
+            TeleFuserVlaStructuredEndpoint,
+            ENDPOINT_METADATA,
+        ),
+        (
+            "transport",
+            "telefuser_structured_http",
+            TeleFuserStructuredHttpTransport,
+            TRANSPORT_METADATA,
+        ),
+    )
+    for category, name, plugin_class, metadata in definitions:
+        if plugins.has_entry(category, name):
+            if not replace:
+                continue
+            plugins.unregister(category, name)
+        plugins.register(category, name, plugin_class, metadata=metadata)
+
+
+__all__ = [
+    "SGLangRealtimeAdapter",
+    "TeleFuserLiveKitAdapter",
+    "TeleFuserStructuredHttpTransport",
+    "TeleFuserVlaStructuredEndpoint",
+    "register_adapters",
+    "register_plugins",
+]

benchmarks/telefuser_aiperf/telefuser_aiperf/cli.py

@@ -6,9 +6,10 @@
 def main() -> None:
     """Register TeleFuser adapters and delegate to the AIPerf CLI."""
 
-    from telefuser_aiperf import register_adapters
+    from telefuser_aiperf import register_adapters, register_plugins
 
     register_adapters()
+    register_plugins()
 
     from aiperf.cli import app