quant_eval_efficiency_and_footprint.csv

run_id,model_id,baseline_kind,quant_type,baseline_bytes,quantized_bytes,compression_ratio,size_reduction_fraction,measurement_scope,excludes,baseline_runner,quantized_runner,execution_substrate,baseline_accelerator,quantized_accelerator,hardware_matched_pair,baseline_tokens_per_second,quantized_tokens_per_second,observed_wall_time_ratio,wall_time_direction,throughput_ratio,token_volume_cost_proxy_ratio,monetary_cost_ratio_status,artifact_footprint_status,hardware_backend_scope
Mistral_Nemo_Instruct_2407_20260814_030505,mistralai/Mistral-Nemo-Instruct-2407,gguf_f16,gguf_q4_k_m,24504279808,7477207744,3.277196601587428,0.6948611506811602,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_llama_cpp,quantized_llama_cpp,local_llama_cpp,nvidia_geforce_rtx_4090,nvidia_geforce_rtx_4090,true,25.043651118780318,91.02962261885648,2.7038164707866055,speedup,3.634838314393905,,suppressed_missing_per_runner_cost_usd,computed,both lanes executed on NVIDIA GeForce RTX 4090; ratio is a precision comparison
Mistral_Nemo_Instruct_2407_20260815_113254,mistralai/Mistral-Nemo-Instruct-2407,gguf_f16,gguf_q5_k_m,24504279808,8727634688,2.8076656143378673,0.643832230272254,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_llama_cpp,quantized_llama_cpp,local_llama_cpp,nvidia_geforce_rtx_4090,nvidia_geforce_rtx_4090,true,30.135280786823675,79.96792404664095,2.564462769097187,speedup,2.653631290590332,,suppressed_missing_per_runner_cost_usd,computed,both lanes executed on NVIDIA GeForce RTX 4090; ratio is a precision comparison
Mistral_Nemo_Instruct_2407_20260816_084553,mistralai/Mistral-Nemo-Instruct-2407,gguf_f16,gguf_q8_0,24504279808,13022372608,1.8817062409154497,0.4685674212817086,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_llama_cpp,quantized_llama_cpp,local_llama_cpp,nvidia_geforce_rtx_4090,nvidia_geforce_rtx_4090,true,33.11406810059508,58.860127886525646,1.7144120311211424,speedup,1.7774961296726903,,suppressed_missing_per_runner_cost_usd,computed,both lanes executed on NVIDIA GeForce RTX 4090; ratio is a precision comparison
Qwen2.5_14B_Instruct_1M_20260815_220633,Qwen/Qwen2.5-14B-Instruct-1M,modal_f16,modal_q4_k_m,29547716384,8988110624,3.287422420581013,0.6958103121340695,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_modal_llama_cpp,quantized_modal_llama_cpp,modal_llama_cpp,nvidia_a100_40gb,nvidia_a10g,false,33.06065710210004,31.686447557979236,0.9614674486693676,slowdown,0.9584336893281679,,suppressed_missing_per_runner_cost_usd,computed,"baseline executed on NVIDIA A100-40GB, quantized executed on NVIDIA A10G; ratio reflects an accelerator change and is not a precision comparison"
Qwen2.5_32B_Instruct_20260815_081051,Qwen/Qwen2.5-32B-Instruct,modal_f16,modal_q4_k_m,65535969920,19851336320,3.3013379484167644,0.6970925074545689,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_modal_llama_cpp,quantized_modal_llama_cpp,modal_llama_cpp,nvidia_a100_80gb,nvidia_a10g,false,18.72691437776273,16.307569717720078,0.8535462010847542,slowdown,0.8708092208230791,,suppressed_missing_per_runner_cost_usd,computed,"baseline executed on NVIDIA A100-80GB, quantized executed on NVIDIA A10G; ratio reflects an accelerator change and is not a precision comparison"
Qwen2.5_7B_Instruct_20260814_234822,Qwen/Qwen2.5-7B-Instruct,gguf_f16,gguf_q4_k_m,15237853632,4683073984,3.2538144142204524,0.6926683969345008,stored_model_weight_artifact_bytes,modal_container_storage;runtime_vram,full_weight_llama_cpp,quantized_llama_cpp,local_llama_cpp,nvidia_geforce_rtx_4090,nvidia_geforce_rtx_4090,true,55.4357001720401,119.12481984641852,1.984891682956756,speedup,2.1488827502263796,,suppressed_missing_per_runner_cost_usd,computed,both lanes executed on NVIDIA GeForce RTX 4090; ratio is a precision comparison