Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
Commit History
18 days ago
-
Fit the picker to the docs iframe Running
The chart drew into a fixed 1180-wide viewBox, so in the ~760px docs iframe its labels shrank to about 6px. It now measures its container and draws at real pixels. Below lg the sentence uses 26px type and tighter padding, and the hover card is clamped inside the chart. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>54513db
30 days ago
-
Recommend methods by source: on the fly or from a checkpoint
The receipt now lists a curated set of methods per precision, in recommendation order, chosen by a new "quantized on the fly / from a checkpoint" slot in the sentence. Adds NVFP4, GGUF, AutoRound and Quark; drops HQQ, optimum-quanto, HIGGS and AQLM, which no list includes. Also drops the ?/* qualifiers from the receipt and the focus outline on the precision hit targets. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>3a22edb
-
Replace the precision dropdown with a hover receipt
The dropdown made you commit to a precision before it would tell you anything. Now the four capabilities that actually decide the choice — can it quantize on load, does it compile, does it fine-tune, does it save — hang off the tick itself. - MethodReceipt lists each method with all four capabilities in fixed lanes, supported in the accent and everything else greyed, so the shape of a row reads before any word does. A legend defines the four once, including "?" for cells the upstream docs leave blank and "*" for partial. - Ticks become 44-unit hit targets with tabIndex and labels: a 2px dotted line is not something you can point at, and it was not reachable by keyboard at all. - Method names link to their Transformers docs page. - Closing waits 160ms so the pointer can travel to the card, which sits outside the container's box and would otherwise dismiss on approach. - The bar opens at the largest precision that fits rather than a default of 4-bit, since there is no long…9b65d3b
-
Build the quantization picker
An interactive replacement for the quantization method table in the Transformers docs. The table lists bits; the deciding arithmetic is bytes, and that is where people stall. This does the arithmetic. - lib/data.ts holds the methods and the memory model: weights from params × bits, KV cache from the attention shape, activations as a flat estimate. - lib/hardware.ts is scraped from huggingface.co/hardware, one entry per memory configuration (316 of them) rather than per device, since an RTX 3060 is sold as both 8GB and 12GB. - The model is typed as a Hub repo id and resolved through /api/model, so any model works, not a curated five. - MemoryBar draws the selected precision solid with fp16 behind it as a ghost, so quantizing shows its own effect. - TradeoffMap places the matching methods by when they quantize and whether they serialize — the axes that actually separate them, since every method at a given precision lands at the same footprint. Deploys to a Space via the D…5ae37d4