{ "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" } }