File size: 2,130 Bytes
9118991 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 | 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
|