$ the-wire · showcase
vLLM v0.32.0 lands system-one endpoint; Ollama exposes template_source
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
16 people shipped this
dhiltgen
@dhiltgen
1 cited
am17an
@am17an
1 cited
allozaur
@allozaur
1 cited
khluu
@khluu
1 cited
positive666
@positive666
1 cited
riteme
@riteme
CISC
@CISC
Todd Malsbary
@malsbat
David Cheung
@sawsa307
andakai
@andakai
Sunt-ing
@Sunt-ing
hnyls2002
@hnyls2002
Oasis-Git
@Oasis-Git
vincentzed
@vincentzed
alphabetc1
@alphabetc1
mickqian
@mickqian
vLLM's v0.32.0 release carries a breaking, default-on structured-decision surface plus a new hf parser pair, while Ollama's API finally reports which template actually renders a model's chat prompts.
v0.32.0 vllm-project/vllm
The 814-commit release adds a POST /v1/systemone endpoint answering choice, noul and score questions over structured state, and serves it by default. It also ships an hf tool and reasoning parser (--tool-call-parser hf, --reasoning-parser hf), so treat this as a breaking release: existing tool and reasoning parser config will need review before you rebuild images on it.
api: populate show renderer/parser fields and add template_source ollama/ollama
ShowResponse's renderer and parser fields existed but were never populated; they now carry effective values, with renderer resolving the gemma4 legacy alias by model size. The new template_source tells you whether the configured renderer/parser, a Go template layer, the GGUF's embedded jinja, harmony, or nothing is rendering chat prompts, and ollama show only prints the rows when set, so output...
ui : Discover & Download Models from HF Hub ggml-org/llama.cpp
The models manager gains a Discover Models view for browsing and downloading HF GGUF models, including featured catalog, debounced search, size-range filters, per-quant download chips with memory-fit notes, and quant-matched draft sidecars that pull down alongside their main quant. It moves model acquisition into the server UI rather than requiring a separate download step.
[Bugfix][Spec Decode] Clear stale shared block-table rows vllm-project/vllm
A larger MTP decode graph could read stale shared block-table rows after the target batch shrank, letting padded work address another request's cache or state. Cleanup now lives in BlockTables, clearing previously active rows during gather and dummy preparation while preserving active rows and independent PCP output buffers; the MTP caller workaround is gone.
CUDA: faster MMVQ for low-bit types ggml-org/llama.cpp
MMVQ picks up kernel speedups for low-bit types, with q4_K and q8_0 sitting at the noise floor as the control, and the gains vary widely by card: 1.71 on A16, 1.59 on RTX 3090, 1.13 on RTX 5090 in the interleaved test-backend-ops run. Also worth noting: llama.cpp added Kolibri-1 support (78B MoE, 3.46B active, bilingual German/English), grouped-experts top-k fusion on CUDA measured 1.20 to 1.21...
Action items
- → Review vLLM tool and reasoning parser config before adopting v0.32.0, which is flagged as a breaking release vllm-project/vllm [plan]
