Maria Liviu Valentin unknownliviu

Qwen3.8-27B on RTX 5090 — 256K, vision, turbo4 KV, fused MMA, MTP

Live stack: Unsloth UD-Q6_K + mmproj-F16, Madreag turbo3-cuda sync/2026-08, native 262144 context, turbo4/turbo4 KV (main + draft), fused MMA + GDN, native MTP n=3.

Hardware this was measured on: RTX 5090 32 GB, CUDA 13.3, SM120.

Perf (RTX 5090, this exact command)

V1 Category Accuracy: ROI Analysis and Options

Ticket: CAT-3897 Context: CAT-3644 accuracy analysis

TL;DR

Recommendation: Do nothing (Option D). The 3.1% V1 L1 error rate translates to ~12K daily broken impressions out of ~378K total (3.1%), but the damage is concentrated in specific patterns and partially absorbed by downstream systems.

Impression impact (from Amplitude ProductCardViewed, 7-day window):

brand_id	brand_name	brand_slug	merchant_slug	allow_listed	reason
569ee4e4-ac73-c052-2104-211545abe8f0	A.N.G.E.L.O. Vintage Cult	a-n-g-e-l-o-vintage-cult	farfetch	TRUE	Legitimate Italian vintage boutique (est. 1978, Lugo). Farfetch partner
eb453d8a-c5a8-cd0a-2f62-b2d5208f862c	A.N.G.E.L.O. Vintage Upcycling	a-n-g-e-l-o-vintage-upcycling	farfetch	TRUE	Legitimate upcycled garment line from A.N.G.E.L.O. Vintage
9608f0a8-ce81-665d-e58c-57a849f7cfdb	Agnona Vintage	agnona-vintage	farfetch	FALSE	Farfetch resale sub-brand: 'Agnona Vintage' follows [Designer] Vintage pattern
88e277b2-647e-f4b6-7087-2049ab96e268	Aigner Vintage	aigner-vintage	farfetch	FALSE	Farfetch resale sub-brand: 'Aigner Vintage' follows [Designer] Vintage pattern
19181cb1-376f-e432-1130-c65add21a8ed	Alaïa Pre-Owned	alaia-pre-owned	farfetch	FALSE	Farfetch resale sub-brand: 'Alaïa Pre-Owned' follows [Designer] Pre-Owned pattern
27e0ec13-0ae2-de74-6d50-355be697950d	Alberta Ferretti Vintage	alberta-ferretti-vintage	farfetch	FALSE	Farfetch resale

	lng	lat
	-84.5017199	34.17301659999999
	-84.4222314	34.1777818
	-84.3319539	34.16692260000001
	-83.18378109999999	42.41683010000001
	-84.47344690000001	45.596342
	-84.3145612	33.8857565
	-81.0943143	32.0797969
	-83.24034890000001	42.4297374
	-83.24041720000001	42.4297384

	console.log('Loading function');
	var http = require("https");
	var fs = require('fs');

	var options = {
	"method": "POST",
	"hostname": "api.sparkpost.com",
	"path": "/api/v1/transmissions?num_rcpt_errors=3",
	"headers": {
	"Content-Type": "application/json",

	// v1.1.2
	var https = require('https');
	var zlib = require('zlib');
	var crypto = require('crypto');

	var endpoint = 'ACTUAL_ENDPOINT'

	// Set this to true if you want to debug why data isn't making it to
	// your Elasticsearch cluster. This will enable logging of failed items
	// to CloudWatch Logs.

	service:
	name: zaius-data-test
	plugins:
	- serverless-step-functions
	- serverless-pseudo-parameters
	provider:
	name: aws
	stage: ${opt:stage, 'sandbox'}
	region: us-east-2
	iamRoleStatements:

	service:
	name: zaius-data
	plugins:
	- serverless-step-functions
	- serverless-pseudo-parameters
	provider:
	name: aws
	stage: ${opt:stage, 'sandbox'}
	region: us-east-2
	resources:

	<link rel="import" href="../core-icon-button/core-icon-button.html">
	<link rel="import" href="../core-toolbar/core-toolbar.html">
	<link rel="import" href="../core-input/core-input.html">
	<link rel="import" href="../core-icons/core-icons.html">
	<link rel="import" href="../core-icon/core-icon.html">

	<polymer-element name="my-element">

	<template>
	<style>

	FROM circleci/ruby:2.5.1-stretch
	RUN sudo apt-get -y -qq update && sudo apt-get install python-pip python-dev build-essential && pip install awscli awsebcli --upgrade --user && echo 'PATH=$PATH:/home/circleci/.local/bin' >> ~/.bashrc