onnx-genai-example-gemma4-e2b / source_provenance.json
justinchuby's picture
Add hidden_states.34 output (folded-carry seed) for the #1716 speculative contract
4b70dd2 verified
Raw
History Blame Contribute Delete
2.16 kB
{
"target": {
"id": "google/gemma-4-E2B-it",
"revision": "3e22461f65e89153144f8adb70e3b8c2cc9845a7",
"license": "apache-2.0",
"license_evidence": "Pinned model card / config: license: apache-2.0",
"source_url": "https://huggingface.co/google/gemma-4-E2B-it/tree/3e22461f65e89153144f8adb70e3b8c2cc9845a7"
},
"build": {
"tool": "mobius",
"mobius_git_sha": "710d4927b1fb6ad68ff0f9589452384b58d99899",
"mobius_git_note": "PR #546 branch justinchuby/gemma4-moe-faithful (pending review, do-not-merge). Adds the gemma4 hidden-state output so a borrowed-KV drafter can seed its folded carry from a real target output. Weights byte-identical to the prior 0776f562 build (model.onnx.data sha256 e525066ebb0284923ddcf12e8e3030a8cf49beb0757a35c06f373256929577eb); only the graph gained the hidden_states.34 output tap.",
"dtype": "float16",
"execution_provider": "cuda",
"features": [
"text-only"
],
"outputs_note": "hidden_states.34 = post-final-norm hidden (lm_head input == HF hidden_states[-1], L4 cos 0.999999); consumed by the paired assistant as onnx-genai #1716 folded_carry_seed. Standalone generation ignores it.",
"optimize": [
"group_query_attention",
"packed_attention",
"skip_norm"
],
"runtime_config": "onnx-genai (inference_metadata.yaml + policies/)",
"canonical_metadata": "inference_metadata.yaml = faithful onnx-genai v1 (validated vs #1716 schema); inference_metadata.mobius.yaml = raw mobius emitter (schema 1.0)"
},
"architecture_contracts": {
"hybrid_attention": "sliding_window(512) + full_attention every 5th layer (types in metadata cache groups)",
"num_hidden_layers": 35,
"num_kv_shared_layers": 20,
"owned_kv_layers": 15,
"heterogeneous_head_dim": {
"local_sliding": 256,
"global_full": 512
},
"num_attention_heads": 8,
"num_key_value_heads": 1,
"per_layer_input_hidden": 256,
"double_wide_mlp": true,
"final_logit_softcapping": 30.0,
"tie_word_embeddings": true,
"moe": "disabled in this checkpoint (enable_moe_block=false); dense MLP \u2014 not invented in metadata"
}
}