[UPDATE][llamacppllmbasev3][1.2.38] Update CUDA engine to b10969

[harveyff]

App Title

llama.cpp Engine Base

Description

Based on beclab/apps main (Chart 1.2.37), update the main llama.cpp image from beclab/ggml-org-llama.cpp:server-cuda12-b10588 to beclab/ggml-org-llama.cpp:server-cuda12-b10969.

Bump Chart/manifest version to 1.2.38, appVersion/versionName to b10969, and refresh the existing English/Chinese upgrade notes. The only workload change is the main container image tag. All other configuration matches the public-source baseline, including llm-init v1.5.0.

Upstream engine highlights (b10588 to b10969)

Chart 1.2.38 — Upgrade llama.cpp from server-cuda12-b10588 to server-cuda12-b10969.

Model compatibility

• Add Qwen3.8-Flash-Next (qwen4exp) text/vision architecture support (#27742), with subsequent recurrent-state rollback and inference fixes.
• Add NVIDIA Nemotron-3-Puzzle-75B-A9B (#25444) and Tencent Hy 4 preview architecture support (#28127). These require matching model weights and sufficient resources.

Speculative decoding and CUDA

• Add DFlash2 draft-model support (#27342); requires a compatible draft checkpoint. Upstream notes that DFlash2 GGUF files converted before August 27, 2026 need reconversion for vision correctness.
• Add CUDA sparse Flash Attention for DeepSeek-V4/GLM (#27970), extend MoE fusion to speculative decoding (#27621), and optimize Q4_K/Q5_K decoding with DGX Spark prefetch improvements (#26705). Gains depend on model and hardware.

Server and chat behavior

• Add --kv-unified-per-slot to cap each request slot when using a unified KV cache; -c controls the total pool (#24124). This update does not change the configured context size automatically.
• Enable preserve_reasoning by default for chat templates that support it; use --no-reasoning-preserve to disable (#28174).
• Accept data: URLs in input_audio and input_video requests, fixing media decoding errors (#27735); requires a compatible multimodal model and client.

Stability and memory

• Reduce peak RAM during model loading when tensor repacking is enabled (#27483).
• Fix the model-router LRU hang with concurrent requests (#28539), speculative decoding after image input (#28715), and CUDA Flash Attention synchronization (#27870).

These are upstream engine changes; model-specific capabilities are not all validated by this chart update. llm-init remains at v1.5.0.
Full upstream changes: ggml-org/llama.cpp@b10588...b10969

Validation

• Final chart directory lint and git diff --check passed.
• CUDA tag availability verified for linux/amd64 and linux/arm64.
• Image b10969 was previously exercised on olarestest001 (Olares 1.12.6, amd64), using a separate test-market-based chart: engine startup at 32768 context and a completion with 6201 prompt tokens succeeded.
• That runtime result validates the image in the test setup, not this exact public-baseline chart. Deployment and historical-version upgrades of this final 1.2.38 chart remain untested.

Statement

• This exact final chart has completed deployment and upgrade validation on the declared Olares version.

[Terminus-Gitbot]

After the modification is completed, please click ready for review to submit the PR

[Terminus-Gitbot]

Chart invalid. Error message: [inconsistent info. Version must be the same in chart. version in OlaresManifest.yaml:1.2.38, Chart.yaml:1.2.38, title:1.2.47]

[Terminus-Gitbot]

After the modification is completed, please click ready for review to submit the PR

[Terminus-Gitbot]

After the modification is completed, please click ready for review to submit the PR

[Terminus-Gitbot]

Check passed, please wait for auto-merge.