Spaces:
Running on Zero
Running on Zero
File size: 895 Bytes
695f3b6 bc3f111 695f3b6 bc3f111 695f3b6 bc3f111 695f3b6 bc3f111 1a6347c 7973967 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 | ---
title: Embedding Quantization CUDA Control
emoji: 🔬
colorFrom: blue
colorTo: yellow
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
license: apache-2.0
---
# Embedding Quantization CUDA Control
A bounded ZeroGPU control lane for the matched local MLX embedding
quantization experiment. The first milestone reproduces the frozen
Qwen3-Embedding-0.6B BF16 vectors on CUDA and compares them with the saved MLX
BF16 vectors. CUDA results do not reproduce MLX/Metal performance.
The second milestone adds CUDA-native bitsandbytes INT8 and NF4 controls.
They are deliberately reported as separate quantizers and are not treated as
equivalents of MLX Q, oQ, or oQe formats.
Matched BF16 controls are also available for GTE-Qwen2-1.5B and
Qwen3-Embedding-8B. Their CUDA vectors are compared against the corresponding
saved MLX BF16 vectors from the private reproduction bucket.
|