$ the-wire · showcase
Ollama preps Phonon voice input, vLLM drops deprecated IR knob
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
18 people shipped this
jmorganca
@jmorganca
2 cited
Woosuk Kwon
@WoosukKwon
1 cited
zhendonghua
@zhendonghua
1 cited
jagrit06
@jagrit06
1 cited
kinetrix
@kinetrix
1 cited
hoyyeva
@hoyyeva
dhiltgen
@dhiltgen
SynapShift
@SynapShift
Sigbjørn Skjæret
@CISC
max-krasnyansky
@max-krasnyansky
allozaur
@allozaur
Ruben Ortlam
@0cc4m
Juan Pérez de Algaba
@jperezdealgaba
Kevin H. Luu
@khluu
Wentao Ye
@yewentao256
BugenZhao
@BugenZhao
AgainstEntropy
@AgainstEntropy
mickqian
@mickqian
Ollama is assembling the pieces for speech-capable local models while vLLM starts retiring vLLM IR and SGLang hands DeepSeek-V4 prompting to DeepSeek's own encoder library.
voice: add voice catalog for Phonon ollama/ollama
`ollama create` now imports `voices/NAME.safetensors` from the checkpoint folder and `ollama show` lists the voices, so a Phonon checkpoint carries its named voices instead of one anonymous default. Voice names match case-insensitively, and omitting one falls back to `default`.
mlxrunner: spell out symbols, numbers and links for Phonon ollama/ollama
Phonon's tokenizer only has pieces for 11 ASCII symbols and no capital accented letters, so everything else reached the model as raw bytes. This ports Gradium's xn-ptts English normalization (the 0.4.1 defaults) and fixes its readings, e.g. `$4.59` becomes "4 dollars and 59 cents" and `$20/month` becomes "20 dollars per month".
[vLLM IR] Deprecate ir_op_priority ahead of vLLM IR removal (#63371) vllm-project/vllm
`ir_op_priority` is deprecated ahead of vLLM IR's removal. If you set it in a config or a plugin, start planning the replacement now rather than waiting for the removal to land.
[Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipe sgl-project/sglang
DeepSeek-V4 and V4.1 chat prompts now render and tokenize through deepseek-recipe 0.2.0, DeepSeek's official encoder, via the `dsv4` / `dsv41` chat encoding adapter; SGLang's own `encoding_dsv4.py` and `encoding_dsv41.py` are deleted. The adapter translates requests on the way in, mapping `developer` to `system`, tool-call argument objects to JSON, and `continue_final_message`.
Adding sliding window attention ml-explore/mlx
Sliding window attention landed in `fast::sdpa`, with Metal kernels for the vector (one and two pass) and matrix (nax and pre-nax) paths plus tests and fallbacks. The PR says benchmarks are to follow, so treat the throughput picture as unsettled.
Fix FFT read/write past buffer end in partial threadgroups ml-explore/mlx
Across the rest: llama.cpp cleared all 33 `npm audit` findings in `tools/ui` and added a behavior-check CI workflow, Vulkan gained ggml src1 precision handling, and MLX fixed an FFT kernel that read and wrote past buffer ends when batch size was not a multiple of `grid.y` (result stayed correct, memory access did not).
Action items
- → Check for `ir_op_priority` usage and move off it before vLLM IR is removed vllm-project/vllm [plan]
