DeepStream SOP Inference Microservice Skill

This skill guides AI coding assistants in building, extending, and debugging the NVIDIA DeepStream SOP (Standard Operating Procedure) Inference Microservice — a GPU-accelerated pipeline for temporal action detection and VLM-based SOP compliance monitoring on industrial video feeds.

Reference repository: https://github.com/NVIDIA/sop-monitoring-blueprints/tree/main/microservices/sop-inference-bp Local reference code: sop-inference-bp/ directory (from a local clone of the repository)

Models

Model-agnostic at both inference stages — swap via env var (and Triton dir for GEBD).

"GEBD" = swappable Stage-1 slot; "DDM" = the default architecture (terms used interchangeably).

Chunking is selectable per request (§ 2): default ddm-net uses GEBD; uniform produces fixed-length chunks and bypasses Stage-1 GEBD (§ 3, § 6). DDM temporal window is configurable via FRAMES_PER_SIDE / SEQUENCE_BATCH (§ 4, § 5), with optional TensorRT (§ 5).

Architecture Overview

Runs in a Docker container (nvds-action-sop) alongside a Kafka container. Full diagram: references/sop_architecture.svg.

Data flow through the 4-stage SOPVideoProcessor pipeline (per-request):

Input Sources                    Docker Container: nvds-action-sop
─────────────                    ──────────────────────────────────────────────────
Video Files ──┐                  FastAPI Server (port 8300)
RTSP Streams ─┤── base64/       ├─ /v1/chat/completions → SOPProcessManager
Basler Camera ┘   file/rtsp/       │
                  camera           │ ModelInitializer: VLM first, then DDM dummy pipeline
                                   │ 4 Thread Pools: cv(32), clip(32), vlm(64), vlm_req(64)
                                   │
                                   ▼ SOPVideoProcessor (per-request)
                                   ┌────────────────────────────────────────────────┐
                                   │ Stage 1: DeepStream Pipeline (GPU)             │
                                   │   Source → nvstreammux → tee1                  │
                                   │    ├─[inference] queue1 → nvdspreprocess       │
                                   │    │  → nvinferserver (Triton CAPI + DDM)      │
                                   │    │  → InferOutputTensorParser → score_queue  │
                                   │    ├─[frames]  queue3 → nvvideoconvert         │
                                   │    │  → capsfilter → appsink                   │
                                   │    │  → DecodedFrameRetriever → frame_queue    │
                                   │    └─[RTSP out] queue → convert → H.264 enc    │  (optional, § 18)
                                   │       → rtppay → udpsink → RTSPServer (§ 18)   │  opt-in only
                                   │              │ boundary scores                 │
                                   │              ▼                                 │
                                   │ Stage 2: Clip Post-Process                     │
                                   │   Boundary detection → chunk segmentation      │
                                   │              │ video frames + timestamps        │
                                   │              ▼                                 │
                                   │ Stage 3: VLM Inference                         │
                                   │   Embedded vLLM (Cosmos Reason 1/2)            │
                                   │   Frame sampling at VLM_FPS → classification   │
                                   │              │ action labels                    │
                                   │              ▼                                 │
                                   │ Stage 4: SOP Checker                           │
                                   │   Sequence validation → missing/misordered     │
                                   │              │ chunk results                    │
                                   │              ▼                                 │
                                   │         final_queue                            │
                                   └────────────────────────────────────────────────┘
                                          │
Output                                    ▼
──────                             ┌─────────────────┐
SSE Stream (chat.completion.chunk) │ Kafka Messages   │
Non-streaming (chat.completion)    │ (JSON/Protobuf)  │
Prometheus metrics (/v1/metrics)   └────────┬────────┘
                                            ▼
                                   Docker Container: kafka
                                   (apache/kafka:3.7.0)

Section Index

Each section is a standalone file in references/ — load only what your task needs.

For end-to-end evaluation, read § 12 first; load build/test/curl/latency/camera/Kafka as needed.

§ 18 is opt-in — generate only when the user explicitly requests RTSP output; otherwise skip § 18 and the RTSP_* rules below.

Key Files Map

The full source-to-target file mapping lives in skill_14_implementation_checklist.md:

• Files copied verbatim from references/ (non-trivial algorithms — cycle detection, qwen_vl_utils preprocessing, DeepStream IOptions API, protobuf sources) with the rationale per file.
• Files copied as adaptable templates (Dockerfile, compose.yaml, Triton config and model.py, ddm_net.py, Pylon emulation config, etc.).
• Files generated from skill sections — each annotated with the Critical Rules below that the generation must follow exactly.
• Docker build prerequisites and post-build verification checklist.

Config files (nvds_preprocess_template.txt, nvds_inference_template.txt, vlm_prompts.txt) are used as-is from configs/.

When skill_06b is loaded, read configs/actions.json from the project root and run the § 6b-G generation workflow to produce nvds_action_detector/missing_number_detector.py. If configs/actions.json is absent or invalid, fall back to copying the reference file.

Critical Rules

Each rule's full detail lives in the linked skill_NN_*.md reference file.

The four RTSP_* rules below apply only when the optional RTSP streaming-output feature (§ 18) is requested. They do not apply to the default build — skip them if the user did not ask for RTSP output.

| RTSP_OUTPUT_TAPS_TEE1 | RTSP output branch links from the existing tee1 (added after the main inference link) only when rtsp_port is present. | skill_18_rtsp_streaming_output.md | | RTSP_LEAKY_QUEUE_TINY | RTSP branch queue must be leaky=2 + tiny cap (max-size-buffers=2) to prevent backpressure and NVMM pool exhaustion. | skill_18_rtsp_streaming_output.md | | RTSP_KEYINT_MAX_30 | RTSP H.264 encoder must set key-int-max=30 (and B-frames disabled) to allow downstream seeking. | skill_18_rtsp_streaming_output.md | | RTSP_ENCODER_FALLBACK | Select software/hardware H.264 encoder based on SW_ENCODER with MJPEG fallback. | skill_18_rtsp_streaming_output.md |