— raw parameter-weight floor
Public checkpoint evidence / no model download
Paste a model repository. Get the honest weight floor.
This Space reads the repository’s public safetensors parameter total and exposes the arithmetic behind INT4, INT8, and BF16 memory planning.
Repository lookup
Identify the exact checkpoint.
Accepts owner/model or a full huggingface.co model URL.
Source state
WAITING FOR MODEL ID
Public metadata receipt
Model weight plan
PARAMETERS×PRECISION BITS÷8 BYTES÷2³⁰×1.20 HEADROOM
— raw parameter-weight floor
— raw parameter-weight floor
Boundary: excludes KV cache, context length, activations, graph capture, allocator fragmentation, quantization metadata, offload, runtime compatibility, speed, quality, safety, and license interpretation.
Carry the evidence forward
Put the source path in the model card.
The badge is a passive SVG. Its destination reloads the exact model lookup on ResearchAudio with a distinct Hugging Face Space attribution source.
For repeatable model releases
Run the same receipt in GitHub Actions.
The versioned action writes this evidence into the workflow summary and emits badge markup attributed to the exact installing repository. It does not modify repository files.
- uses: deepmehta11/researchaudio-hf-vram-action@v1
with:
model: owner/modelDecision sequence
A weight floor is the first line, not the verdict.
- 01
Confirm identity. Repository, revision, and exact artifact still matter.
- 02
Read the floor. Parameter arithmetic exposes the minimum resident-weight order of magnitude.
- 03
Add context. KV cache grows with architecture, sequence length, concurrency, and precision.
- 04
Measure runtime. Kernel support, throughput, and real allocations decide deployment.