RFD 1027 details: the tier table, staging, quantization, what to do
Tier per model
A tier must hold the weights and the activations together. The Peak column is the resident weight set, which is smaller than the total for a model that runs in stages.
| Model | Weights | Peak | Tier |
|---|---|---|---|
| qwen_q4_k_m_image_edit | 14.85 GB | 14.85 GB | 24 GB |
| pixal3d_image_to_textured_mesh | 24.05 GB | 6.50 GB | 24 GB |
| krea2_turbo_text_to_image | 9.30 GB | 9.30 GB | 24 GB |
| seethrough_layer_decomposition | 9.82 GB | 5.13 GB | 24 GB |
| trellis2_image_to_textured_mesh | 8.00 GB | 8.00 GB | 24 GB |
| worldmirror2_reconstruct | 2.40 GB | 2.40 GB | 24 GB |
| triposplat_image_to_splat | 2.20 GB | 2.20 GB | 24 GB |
| skintokens_auto_rig | 1.00 GB | 1.00 GB | 24 GB |
| p3sam_mesh_segmentation | 0.80 GB | 0.80 GB | 24 GB |
| kimodo_text_to_motion | 0.60 GB | 0.60 GB | 24 GB |
Every model reaches a 24 GB card. That is the finding, and it is the opposite of what this RFD first recorded.
Staging is what makes the tier
Pixal3D holds 24.05 GB and peaks at 6.50 GB. Three stages run in order, and each frees before the next loads. Without that staging it would need an 80 GB card.
See-Through holds 9.82 GB and peaks at 5.13 GB, for the same reason. RFD 1044 makes the load and the unload real actions, thus the peak is a planning result.
A model image that loads every stage at once pays for a larger card every second it runs.
Quantization is now a cost choice
On one shared device, quantization decided what fit. On Replicate it decides what a second costs.
Qwen at 54.0 GB in bf16 needs an 80 GB card. At 14.85 GB in Q4_K_M it needs 24 GB. Both run, and the tiers differ in price.
Measure the quality cost before the catalog depends on it. RFD 1043 records that no such measurement exists yet.
What to do
- Size each model alone. A set total means nothing here.
- Stage the loads, and report the peak and not the sum.
- Prefer bf16 over fp16 where a checkpoint offers both. The memory is the same, and bf16 carries the wider exponent range.
- Add the activation peak for the resolution and the batch size of each job. The tiers above count weights only.