fix README inconsistency v1

Add 2x24GB and 2x32GB builds, and rename the build directories by card size

Announce the 4x3090 quant in the README

Add a 4x3090 compose file for the main weights

Fix the 4x3090 build for the 4x3090 weights

Point 4x3090 users to the 4x3090 branch v2

Point 4x3090 users to the 4x3090 branch

Credit the author of the upstream vLLM fix

Fix looping output and engine crash with MTP when two requests run at once

Update README.md

Fix structured output failing with MTP when two requests run at once

Keep the vLLM and Triton caches in ./cache

rewrite the docs, compose.yaml, Dockerfile, and add image and MTP support for 3x 3090

Keep recently read QSA K/V rows in VRAM during decode

Add benchmark evaluations

some decode optimizations

Pin the QSA host K/V pool exactly and check it against free RAM

initial commit