v0.8.35 — install.sh resolves vLLM 0.32.0 again; Kolibri-1 DSpark drafter

What's new

• install.sh pins vLLM 0.32.0, gradio>=6 and transformers 5.19.0 (#188). This fixes the known issue in the 0.8.34 notes. vLLM 0.33.0 requires pillow>=13, and every gradio 6 release caps pillow<13. So a fresh install with the default components (core,vllm,chat) had backtracked to vLLM 0.29.0 and gradio 3.36.1, and glq-chat cannot run on that pair. With the pins, the same install resolves without backtracking. The manual pip install vllm==0.32.0 … workaround from the 0.8.34 notes is no longer needed.
• glq-code / glq-chat: --dspark knows Kolibri-1's drafter, intuit/Kolibri-1-dspark (#187). It loads without remote code and serves through the aleph-alpha-inference plugin's vLLM (0.29.x).
• The distro matrix now smoke-tests gemma-4-E4B through the chat template, because the SmolLM3 checkpoint it used is gone from the Hub.
• README: the Kolibri-1 DSpark row now needs glq ≥ 0.8.35 instead of "glq from main".

Validation

• Tests (3.12 / 3.13 / 3.14) green on #187, #188 and the release PR (#189).
• #188: distro matrix in source mode; ubuntu-source and fedora43-source passed on an RTX PRO 6000 Blackwell. A chat-component install in a fresh ubuntu:24.04 container exited 0 in 282 s with no pip backtracking, and pip check was clean.
• #187: Kolibri-1 + DSpark went from 53.31 to 118.56 tok/s (2.22×), measured at B=1 on an RTX PRO 4500 with vLLM 0.29 and the plugin.

Existing checkpoints

Unaffected: no quantize-path or kernel changes.