$ the-wire · showcase
llama.cpp 0.4.1 lands Maple 20B-A1B, Ollama drops typical_p
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
15 people shipped this
dhiltgen
@dhiltgen
1 cited
Georgi Gerganov
@ggerganov
1 cited
wuhangxian
@wuhangxian
1 cited
liranschour
@liranschour
1 cited
sherlockwu
@sherlockwu
1 cited
ParthSareen
@ParthSareen
github-actions[bot]
@github-actions[bot]
Alex
@AlexGabbia
mkhazraee
@mkhazraee
Kevin H. Luu
@khluu
Nick Hill
@njhill
windscope
@windscope
amd-danli103
@amd-danli103
ichbinblau
@ichbinblau
pllimax
@pllimax
llama.cpp's 0.4.1 release adds three new model architectures and a heap-corruption fix, while Ollama's create pipeline shifts weight conversion out of tree and retires the typical_p sampling parameter.
ggml-cpu : disable PCH and fix CACHE_LINE_SIZE ambiguity to fix heap corruption (#28882) ggml-org/llama.cpp
A precompiled header force-included ggml-impl.h before ops.h, which pulled in <new> and set CACHE_LINE_SIZE to the 256-byte hardware destructive interference size in the C++ kernels while the C work-buffer sizing code used a different value. Removing the PCH and the std::hardware_destructive_interference_size branch resolves the heap corruption that mismatch caused. If you build ggml-cpu yourse...
create: add server-side MLX imports and drop GGUF conversion ollama/ollama
Safetensors imports now run through the MLX create pipeline locally and on the server, with remote upload/staging, draft layer handling, cancellation propagation, and transfer limits. The in-tree safetensors-to-GGUF converter, server quantization path, and converter-only dependencies are gone: GGUF conversion and quantization now live in llama.cpp tooling, and GGUF create is limited to wrapping...
[KVConnector][P2P] Configurable unbound-store timeout and one-RTT rejection of a late fetch vllm-project/vllm
In the p2p secondary offloading tier, a producer parks a request's stored KV blocks until the consumer's FetchMsg binds them to a session, then reaps them after 60 s so they stop pinning primary-tier CPU slots. That 60 s ceiling is now the unbound_store_timeout_s tier config key, defaulting to the existing value, which gives deployments with prefills that legitimately run longer a way to raise it.
[Bugfix] Handle bare and malformed tool call openers in Gemma4 parser vllm-project/vllm
gemma-4-26b-a4b-it emits tool calls in formats the parser's FSM had no transition for: a bare <|tool_call>:name{...} opener, an opener with no brace, and round-bracket arguments. The tool call was silently dropped in the first case. This is the kind of failure that shows up as a missing tool result rather than an error, so it is worth re-testing against 0.4.1 if you run Gemma4 tool calling.
[Rust] Use Dynamo native renderers when chat templates are missing sgl-project/sglang
Models such as DeepSeek V4 ship no chat_template, and the Rust frontend had responded by disabling /v1/chat/completions for them entirely. When tokenizer_config.json has no template and no --chat-template is passed, it now falls back to dynamo_renderer::native_formatter_for, with Python sending the HF model_type for selection and dynamo-renderer bumped to 5.1.2.
Action items
