| model: | |
| repository: "tomg-group-umd/huginn-0125" | |
| revision: "bb6621b65e90b6a4b9b29ef88dc83866d450470c" | |
| config_sha256: "e9fe79df06a783ca33a76038c59a715b6a62f15c4a5b17b9681e697fea46c79c" | |
| modeling_sha256: "a1d447da93c605a6dc11fbc17cbc7185663f4a41ae12335b34b5847edf0f84aa" | |
| physical_layers: 4 | |
| recurrent_loops: 32 | |
| hidden_size: 5280 | |
| intermediate_size: 17920 | |
| groups: | |
| attention_qkv: | |
| hf_weights: [attn.Wqkv] | |
| vllm_consumer: self_attn.qkv_proj | |
| attention_output: | |
| hf_weights: [attn.proj] | |
| vllm_consumer: self_attn.o_proj | |
| mlp_up_gate: | |
| hf_weights: [mlp.fc] | |
| vllm_consumer: mlp.gate_up_proj | |
| mlp_down: | |
| hf_weights: [mlp.proj] | |
| vllm_consumer: mlp.down_proj | |
| slt_budget: | |
| rule: "ceil(5% of candidate transform parameters), then select whole groups until the target is met" | |
| fixed_group_count: null | |
| status: "requires calibrated transform factorization and candidate parameter counts" | |
| runtime_status: | |
| adapter_contract: implemented | |
| vllm_hook_integration: implemented_cpu_validated | |
| activation_sites: [attention_qkv, attention_output, mlp_up_gate, mlp_down] | |
| weight_dispatch: "shared W4 plus recurrence-indexed variants for selected SLT groups" | |
| cta_boundaries: 31 | |
| method_isolation: "reject NVFP4, ResidualQuant, MomentumQuant, and residual activation policies" | |
| artifact_export: >- | |
| Implemented with a mandatory progressive SLT audit using the 5% parameter | |
| target and whole-group overshoot rule; arbitrary fixed group counts are rejected. | |
| calibration: implemented_cpu_validated_gpu_not_run | |
| calibration_command: >- | |
| CUDA_VISIBLE_DEVICES=0 | |
| PYTHONPATH=./loopQ | |
| python loopQ/scripts/calibrate_huginn.py | |
| --gpu 0 --optimizer adamw --learning-rate 5e-3 --total-steps 1500 | |
| --slt-round-steps 25 --hidden-factors 60,88 | |
| --intermediate-factors 128,140 | |
| --output ${LOOPTRANSFORMER_DATA_ROOT}/checkpoints/loopq/huginn_w4a4_components.pt | |
| smoke_generation: not_run | |
| smoke_command: >- | |
| CUDA_VISIBLE_DEVICES=0 | |
| PYTHONPATH=./loopQ | |
| python loopQ/scripts/smoke_huginn_vllm.py | |
| --artifact ${LOOPTRANSFORMER_DATA_ROOT}/models/loopq/huginn_w4a4.pt | |
| --gpu 0 | |