{
  "model_type": "fela",
  "library_name": "pytorch",
  "arch": "CPUGPT (FELA LM)",
  "edition": "FELA LLM 1.5 (code, 1.5B class)",
  "note": "FELA code language model for fill in the middle autocomplete. 1.79B parameters. 28 layers of FNO spectral token mixers plus Gated-DeltaNet recall plus Landmark global routing (SSSL pattern). Final fill in the middle SFT model.",
  "quant": "bf16",
  "vocab_size": 151936,
  "n_layer": 28,
  "n_embd": 1536,
  "n_head": 12,
  "ffn_hidden": 8960,
  "fno_modes": 512,
  "gla_chunk": 256,
  "layer_pattern": "SSSL",
  "gla_delta": true,
  "landmark_layer_every": 7,
  "landmark_chunk": 32,
  "landmark_max": 64,
  "attn_layer_every": 0,
  "seq_len": 2048,
  "n_params": 1788379536,
  "tokenizer": "Qwen2.5-Coder (tokenizer.json, vocab 151936, carries the FIM tokens <|fim_prefix|> / <|fim_suffix|> / <|fim_middle|>)",
  "fim_tokens": {
    "prefix": "<|fim_prefix|>",
    "suffix": "<|fim_suffix|>",
    "middle": "<|fim_middle|>"
  },
  "input_desc": "token ids from the Qwen2.5-Coder tokenizer. For fill in the middle: <|fim_prefix|> P <|fim_suffix|> S <|fim_middle|>, and the model completes the middle. For plain continuation: the prompt token ids.",
  "serving": "Runs on a plain CPU with no GPU via the FELA server (int8, constant memory streaming decode, no KV cache). The bf16 model.safetensors is the reference; model_int8.safetensors is the int8 serving export.",
  "int8_variant": "model_int8.safetensors",
  "int8_scheme": "weight only int8 per channel symmetric on linear layers; scales fp32; kept tensors bf16; keys prefixed int8 / scale / keep",
  "architectures": [
    "FelaForCausalLM"
  ],
  "auto_map": {
    "AutoConfig": "configuration_fela.FelaConfig",
    "AutoModelForCausalLM": "modeling_fela.FelaForCausalLM"
  }
}