{
  "schema_version": 1,
  "_comment": "InternVL3.5-1B (InternViT-300M vision + Qwen3-0.6B LLM) image->caption on Hexagon v81 (SM8850 / soc_model 87) via the internvl_generate host-op. 448px image -> 256 image tokens (host patch-embed + pixel-shuffle + MLP projector) spliced into the text prompt, right-padded to the prefill window Pn=512, one batched prefill seeds the decode KV, then AR=1 decode + in-graph lm-head. The recipe's prefill graph (ivl_pf, batched GQA) and lm-head graph (ivl_lmh) were reconstructed (export_llm_prefill.py / export_lmhead_graph.py) — gated vs the HF gold first token. All graphs f16, f16 I/O. Compiled via the DLC route + {O:3,vtcm_mb:8,dlbc:1}.",
  "model": { "name": "internvl3_5-1b", "family": "vlm", "dsp_arch": "v81" },
  "params": {
    "hidden": 1024, "vocab": 151936, "n_layers": 28, "max_ctx": 512,
    "kv_dim": 1024, "head_dim": 128, "rope_theta": 1000000.0,
    "vision_px": 448, "eos_token_id": 151645
  },
  "artifacts": {
    "contexts": {
      "vision":  { "bin": "ivl_enc_f16.bin" },
      "prefill": { "bin": "ivl_pf_f16.bin" },
      "decode":  { "bin": "ivl_dec_f16.bin" },
      "lmhead":  { "bin": "ivl_lmh_f16.bin" }
    },
    "embed": "ivl_embed_f16.bin",
    "tokenizer": "tokenizer.json",
    "fixture_dir": "vlm"
  },
  "plan": {
    "steps": [
      { "host": "internvl_generate",
        "params": { "vision": "ivl_enc_f16", "prefill": "ivl_pf_f16", "decode": "ivl_dec_f16", "lmhead": "ivl_lmh_f16" } },
      { "emit": "all" }
    ]
  }
}