$ the-wire · showcase
MiniCPM-V 4.7 lands in llama.cpp, vLLM fixes KV reservation under pressure
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
Today's wire is mostly plumbing: a new vision model reaches llama.cpp, vLLM tightens block reservation and video cache identity, and sglang-diffusion stops leaving multi-GPU throughput on the floor.
model : support MiniCPM-V 4.7 (#29416) ggml-org/llama.cpp
MiniCPM-V 4.7 is structurally near-identical to 4.6, with 3D RoPE as the one architectural delta; the implementation reuses the existing 4.6 path and extends it, so conversion and loading work through the normal route.
[Bugfix][KV Connector] Fix speculative block reservations across KV cache groups vllm-project/vllm
Speculative decode block reservations for asynchronous KV loads now account for every KV cache group using its own allocation block size. Under KV memory pressure, under-reserving previously left a request admitted for an async load but unable to start decoding once the load finished; the promotion-margin regression test now covers single-group, multi-group, and mixed-block-size configurations.
[diffusion] torch.compile: keep the TP all-reduce inside the graph sgl-project/sglang
With `--enable-torch-compile` and `--tp-size > 1`, each row-parallel layer was breaking the DiT graph at its all-reduce because Dynamo cannot trace the tvm_ffi communicator's `get_world_size`. The fix keeps the TP all-reduce inside the compiled graph.
[Bugfix][Multimodal] Fix per-video cache option reuse vllm-project/vllm
Each Qwen3-VL video's selected `fps` and `num_frames` now contribute to its cache identity, so repeated items and partial multimodal-cache hits reuse the options that actually apply. Option lists that are too short are rejected before indexing rather than silently misaligned.
[diffusion] BCG: warn instead of disabling on unvalidated models sgl-project/sglang
Elsewhere in the diffusion desk: `--enable-breakable-cuda-graph` no longer turns itself off on models absent from the validated list, warning instead, and dynamic LoRA deltas are folded into the second GEMM.
