{
  "mic": "0.3",
  "name": "RecurrentGemma causal LM",
  "usage": "AutoTokenizer + AutoModelForCausalLM, batched causal LM with left padding, inference logits and one SGD step",
  "architecture": "RecurrentGemmaForCausalLM",
  "checkpoints": {
    "google/recurrentgemma-2b": {
      "fixture": "huggingface-mic/recurrentgemma-2b-mic"
    }
  },
  "fixtures": {
    "huggingface-mic/recurrentgemma-2b-mic": {
      "built_from": "google/recurrentgemma-2b",
      "covers": [
        "google/recurrentgemma-2b"
      ],
      "build": {
        "seed": 17,
        "config_overrides": {
          "fields": {
            "num_hidden_layers": 3
          },
          "reason": "The generic rule shrinks to 2 layers, and block_types (recurrent, recurrent, attention) makes both of them recurrent: the fixture would have no attention block, so attention_mask is never read and padding goes untested. A third layer adds one sliding-window attention block, as in the checkpoint's repeating pattern."
        }
      }
    }
  },
  "integrations": {
    "transformers": {
      "entry_point": "integrations/transformers/usage.md",
      "format": "runnable-markdown",
      "requires": [
        "transformers>=5",
        "torch"
      ],
      "targets": [
        "fixtures",
        "checkpoints"
      ],
      "variants": {
        "default": {},
        "eager": {
          "attn_implementation": "eager"
        },
        "bfloat16": {
          "dtype": "bfloat16"
        }
      },
      "training": true,
      "resources": {
        "fixture": {
          "memory_gb": 4,
          "download_mb": 40
        },
        "google/recurrentgemma-2b": {
          "devices": [
            "cuda"
          ],
          "gpu_memory_gb": 18,
          "memory_gb": 16,
          "download_mb": 10750,
          "variants": {
            "bfloat16": {
              "gpu_memory_gb": 10
            }
          }
        }
      }
    }
  },
  "outputs": {
    "runtime": {
      "compare": "exact"
    },
    "logits.last": {
      "compare": "allclose",
      "atol": 0.0001,
      "rtol": 0.0001
    },
    "logits.first_real": {
      "compare": "allclose",
      "atol": 0.0001,
      "rtol": 0.0001
    },
    "training.loss": {
      "compare": "allclose",
      "atol": 0.0001,
      "rtol": 0.0001
    },
    "training.gradients": {
      "compare": "all_true"
    },
    "training.updated": {
      "compare": "true"
    }
  }
}