Skip to main content

FitVerdict

Whether a model fits one device, judged from its metadata: every term of the law and the verdict.

activation_bytes (property)​

The working set priced beside the weights and the cache.

context_length (property)​

The model's own context length; 0 when its documents declare none.

device (property)​

The device's label.

device_bytes (property)​

The device's total memory.

fits (property)​

need_bytes <= device_bytes.

grade (property)​

How comfortably the need sits in the device's memory: 'Perfect' (at most 60% of it), 'Good' (at most 85%), 'Marginal' (at most 98%), 'Too tight' (above, or does not fit); empty when not judged.

judged (property)​

True once the source was priced on the device; False on a row whose source could not be read.

kv_bytes_fixed (property)​

Cache bytes a session holds whatever its length: sliding-window rings, conv slabs, per-row bookkeeping.

kv_bytes_per_token (property)​

Cache bytes per token of context.

kv_dtype (property)​

The cache dtype (bfloat16, float16, float32); empty when no cache is priced.

max_image_fits (property)​

An image generation pipeline: the side of the largest square image that fits; 0 otherwise.

max_seq (property)​

The sequence length judged.

max_seq_fits (property)​

The largest sequence length that fits, at most the context length.

memory_share_percent (property)​

need_bytes over device_bytes as a whole percent, rounded up.

need_bytes (property)​

weight_bytes + kv_bytes_fixed + kv_bytes_per_token * max_seq + activation_bytes.

parameters (property)​

The parameter count the speed estimate divides by: a GGUF header's tensor table, or the weight bytes over the declared dtype's width; 0 when unknown.

reason (property)​

What decided the verdict.

source (property)​

The model source as given.

tok_s_estimate (property)​

The estimated decode speed on a CPU in tokens per second; 0 when no estimate applies.

tok_s_measured (property)​

A measured decode speed for the device; 0 when none is known.

weight_bytes (property)​

The weight files a fetch takes.

weight_encoding (property)​

The weight encoding the speed estimate reads its factor from (a GGUF file type such as 'Q4_K_M', or a declared stream dtype such as 'bfloat16'); empty when unknown.

__init__​

__init__(self, /, *args, **kwargs)

Initialize self. See help(type(self)) for accurate signature.