1.2.1: MTP speculative decoding

docs(cookbook): measured 16 GB recipe + The 16 GB tier, fp8 KV cache, 3 troubleshooting entries

docs: 16 GB tier measured, developer-role 400, sustained-load stability note

Clarify that 64k context and 8-9 streams are not simultaneous

Correct RADIX guidance (prefix cache is exact-match only); add long-context sizing

escha 1.2.0: docs + changelog for tensor parallelism

Remove superseded 1.1.1 wheel (replaced by 1.2.0)

escha 1.2.0 wheel: tensor parallelism (gated on world_size>1)

escha 1.1.1: docs + changelog for the cuda:0 device fix

escha 1.1.1 wheel: rank device instead of hardcoded cuda:0

Remove superseded 1.1.0 wheel (replaced by 1.1.1)

Escha Runtime qwen3dense — escha 1.1.0+qwen3dense SGLang wheel, serve.sh, cookbook