{
  "schema_version": 1,
  "_comment": "InternVL3.5-1B with a PADDED prefill (ivl_pf_512, AR=512) so the TEXT prompt is no longer locked to S=310 — arbitrary prompts up to ~256 text tokens work (image stays fixed at 448px -> 256 tokens). internvl_generate right-pads the image-spliced prompt to Pn and seeds only the real S decode rows.",
  "model": { "name": "internvl3_5-1b-512", "family": "vlm", "dsp_arch": "v79" },
  "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_enc2.bin" },
      "prefill": { "bin": "ivl_pf_512.bin" },
      "decode":  { "bin": "ivl_dec_w4.bin" },
      "lmhead":  { "bin": "ivl_lmh_q.bin" }
    },
    "embed": "ivl_embed_f16.bin",
    "tokenizer": "tokenizer.json",
    "fixture_dir": "vlm"
  },
  "plan": {
    "steps": [
      { "host": "internvl_generate",
        "params": { "vision": "ivl_enc2", "prefill": "ivl_pf_512", "decode": "ivl_dec_w4", "lmhead": "ivl_lmh_q" } },
      { "emit": "all" }
    ]
  }
}