{
  "id": "experiments/debug/gemma3-debug-q4k",
  "name": "gemma3-q4k-debug",
  "description": "Debug Gemma 3 1B Q4K inference - check kernel selection and alignment",
  "intent": "investigate",
  "compatibleIntents": ["investigate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-02-25T00:00:00Z",
  "extends": "default",
  "model": "gemma-3-1b-it-q4k-ehf16-af32",
  "runtime": {
    "shared": {
      "debug": {
        "logLevel": {
          "defaultLogLevel": "verbose"
        },
        "trace": {
          "enabled": true,
          "categories": [
            "kernels",
            "ffn",
            "attn",
            "embed",
            "logits"
          ],
          "layers": [
            0,
            1
          ],
          "maxDecodeSteps": 5
        },
        "probes": [
          {
            "id": "embed_out",
            "stage": "embed_out",
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1150,
              1151
            ]
          },
          {
            "id": "layer0_attn_input",
            "stage": "attn_input",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1150,
              1151
            ]
          },
          {
            "id": "layer0_attn_normed",
            "stage": "attn_normed",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1150,
              1151
            ]
          },
          {
            "id": "layer0_q_proj",
            "stage": "q_proj",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1020,
              1021,
              1022,
              1023
            ]
          },
          {
            "id": "layer0_attn_out",
            "stage": "attn_out",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3
            ]
          },
          {
            "id": "layer0_post_attn",
            "stage": "post_attn",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1150,
              1151
            ]
          },
          {
            "id": "layer0_ffn_in",
            "stage": "ffn_in",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              1150,
              1151
            ]
          },
          {
            "id": "layer0_ffn_out",
            "stage": "ffn_out",
            "layers": [
              0
            ],
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3
            ]
          },
          {
            "id": "logits_final",
            "stage": "logits",
            "tokens": [
              -1
            ],
            "dims": [
              0,
              1,
              2,
              3,
              138,
              3730
            ]
          }
        ],
        "kernelTrace": {
          "layers": [
            0,
            1
          ],
          "breakOnAnomaly": true,
          "explosionThreshold": 10,
          "collapseThreshold": 0.000001,
          "maxSteps": 100
        }
      }
    },
    "inference": {
      "prompt": "Hello",
      "batching": {
        "maxTokens": 8
      },
      "sampling": {
        "temperature": 0,
        "topK": 1,
        "topP": 1
      }
    }
  }
}