feat(config): default prompt_cache_all to true

[localai-bot]

Summary

Make the per-request cache_prompt knob default to on, matching upstream llama.cpp.

backend/cpp/llama-cpp/grpc-server.cpp unconditionally forwards the proto field:

data["cache_prompt"] = predict->promptcacheall();   // grpc-server.cpp:197

Since the YAML-loaded Go value was a plain bool (zero = false), any model that didn't explicitly set prompt_cache_all: true ended up sending cache_prompt=false to llama.cpp — overriding upstream's own default (common/common.h:592: bool cache_prompt = true). With kv_unified=true and cache_idle_slots=true already default in parse_options, this was the last piece keeping the per-request prompt cache from being usable out of the box.

Change

• Make LLMConfig.PromptCacheAll tristate (*bool), mirroring MMap, MMlock, Reranking, etc.
• In SetDefaults, when nil → set to true.
• Dereference at the proto boundary (gRPCPredictOpts is post-SetDefaults, so non-nil by contract — same idiom as the surrounding *c.Temperature, *c.TopK lines).
• Add three Ginkgo specs covering: default, explicit false, explicit true. Mirrors the existing enable_prefix_caching precedent in hooks_test.go.

Notes

• No proto change. Generated pb.PredictOptions.PromptCacheAll stays bool.
• Users can still opt out with prompt_cache_all: false in the model YAML.
• No in-tree YAML currently sets prompt_cache_all: false, so the behavior flip lands as a pure improvement.

Test plan

• go test ./core/config/ ./core/backend/ — 102 specs (up from 99) pass
• go vet ./core/config/ ./core/backend/ clean
• CI lint job (golangci-lint) — verify in PR
• Manual smoke: serve a chat model without prompt_cache_all in YAML, confirm second request with shared prefix is faster than the first

🤖 Generated with Claude Code