{
  "mic": "0.3",
  "name": "SwitchTransformers text-to-text generation",
  "usage": "AutoTokenizer + AutoModelForSeq2SeqLM.generate, batched, with a seeded training step",
  "architecture": "SwitchTransformersForConditionalGeneration",
  "checkpoints": {
    "google/switch-base-8": {
      "fixture": "huggingface-mic/switch-base-8-mic"
    }
  },
  "fixtures": {
    "huggingface-mic/switch-base-8-mic": {
      "built_from": "google/switch-base-8",
      "covers": [
        "google/switch-base-8"
      ],
      "build": {
        "seed": 17,
        "model_class": "SwitchTransformersForConditionalGeneration"
      }
    }
  },
  "integrations": {
    "transformers": {
      "entry_point": "integrations/transformers/usage.md",
      "format": "runnable-markdown",
      "requires": [
        "transformers>=5",
        "torch",
        "sentencepiece"
      ],
      "targets": [
        "fixtures",
        "checkpoints"
      ],
      "variants": {
        "default": {},
        "eager": {
          "attn_implementation": "eager"
        },
        "bfloat16": {
          "dtype": "bfloat16"
        }
      },
      "variants_note": "The checkpoint stores bfloat16 weights; the default recipe loads them as-is. The eager variant forces the eager attention path. The bfloat16 variant pins bfloat16 explicitly.",
      "variants_override": {
        "default": {
          "dtype": "float32"
        },
        "eager": {
          "attn_implementation": "eager",
          "dtype": "float32"
        },
        "bfloat16": {
          "dtype": "bfloat16"
        }
      },
      "training": true,
      "resources": {
        "fixture": {
          "download_mb": 200
        },
        "google/switch-base-8": {
          "memory_gb": 8,
          "download_mb": 2000
        }
      }
    }
  },
  "outputs": {
    "runtime": {
      "compare": "exact"
    },
    "generated_tokens": {
      "compare": "exact"
    },
    "generated_text": {
      "compare": "exact"
    },
    "label_logprobs": {
      "compare": "allclose",
      "atol": 0.0001,
      "rtol": 0.0001
    },
    "decoded_roundtrip": {
      "compare": "exact"
    },
    "training.loss": {
      "compare": "allclose",
      "atol": 0.0001,
      "rtol": 0.0001
    },
    "training.gradients": {
      "compare": "all_true"
    },
    "training.updated": {
      "compare": "true"
    }
  }
}