Spark

Serve forward + HTTP/API

How Spark serves tiny CPU forward from SPARK_BC / safetensors. Spark / SparkLang only.

Tiny CPU serve (file SERVE)

make spark-serve
        ./spark-serve docs/examples/spark-builder.sparkbc /tmp/serve-dry-001

        PYTHONPATH=python python3 tools/spark-bc-dump/dump.py \
         docs/examples/spark-builder.sparkbc \
         --serve /tmp/serve-dry-001 \
         --serve-weights /path/to/weights.safetensors # optional

Writes SERVE with forward=true and trained from weights meta. Forward path when MLP tensors exist:

embed_mean_pool->mlp->rms_norm->lm_head

Code: python/sparklang/model_lab/serve.py. Gate: make test-sparkbc (test_serve_forward).

Architecture: Architecture · Attention / forward.

HTTP / stdio API (serve API path)

make spark-serve-api
        ./spark-serve-api --weights /tmp/serve-dry-001/weights.safetensors \
         --http --host 127.0.0.1 --port 8765

        curl -s http://127.0.0.1:8765/health
        curl -s http://127.0.0.1:8765/version
        curl -s -X POST http://127.0.0.1:8765/v1/predict \
         -H 'content-type: application/json' \
         -d '{"token_ids":[65,66,67]}'
        curl -s -X POST http://127.0.0.1:8765/v1/embeddings \
         -H 'content-type: application/json' \
         -d '{"token_ids":[65,66]}'

        # stdio JSON lines:
        ./spark-serve-api --weights … --stdio

        make test-serve-api
Endpoint Role
GET /health Liveness (production: false)
GET /version Product / version / lane G
POST /v1/predict Next-token via tiny CPU forward
POST /v1/embeddings Embeddings from Spark weights

Implementation: python/sparklang/model_lab/serve_api.py wraps run_tiny_forward / run_tiny_embed — same path as file SERVE (now with attention when tensors exist). Builder §6c: SPARK_BC Builder.

Optional live gateway ask/embed (the AI gateway etc.) is a different surface — AI_MODELS.md / ASK_LIVE.md.

Never

  • Claim production LLM serving.
  • Claim multi-layer / rotary-embedding / KV-cache decode.
  • Claim a competitive AI win from SERVE / predict JSON.