feat(ds4): SSD streaming + quality engine options, 128GB DeepSeek gallery models

[localai-bot]

What

Makes ds4's tunable engine knobs reachable from LocalAI model YAMLs, and adds DeepSeek gallery models targeting the 128 GB class - including the full 4-bit Flash quant that only fits via SSD streaming.

Why

The ds4 backend zero-initialized ds4_engine_options and wired none of the engine's options, so SSD streaming (run a model larger than RAM by streaming routed MoE experts from the GGUF on SSD) and the quality/perf knobs were unreachable. SSD streaming is exactly what turns "does the model fit in RAM" into a speed spectrum, so the 153 GB Flash quant can run on a 128 GB machine.

Backend

LoadModel now maps ModelOptions.Options[] onto ds4_engine_options through a declarative table (kEngineOptSpecs + apply_engine_option) rather than per-field branches. ds4_engine_options is fixed C with no reflection, so the field set is enumerated once; adding a future knob is a one-line table row. Details:

• Two fields use ds4's own typed parsers - ssd_streaming_cache_experts (count or NGB, sets experts+bytes via ds4_parse_streaming_cache_experts_arg) and simulate_used_memory (NGB via ds4_parse_gib_arg).
• Bare flags (e.g. ssd_streaming with no value) mean true.
• Path-type options (mtp_path, expert_profile_path, directional_steering_file) resolve relative to the model directory, so a gallery entry can reference a companion file by bare filename; absolute values pass through.
• mtp_draft / mtp_margin are now validated instead of parsed with throwing std::stoi/std::stof.
• ds4_role / ds4_layers / ds4_listen / ds4_route_timeout / kv_cache_dir keep their dedicated handling (validation + coordinator wiring).

Newly reachable keys: ssd_streaming, ssd_streaming_cold, ssd_streaming_cache_experts, ssd_streaming_preload_experts, simulate_used_memory, warm_weights, quality, power_percent, prefill_chunk, expert_profile_path, directional_steering_file/attn/ffn.

Gallery (128 GB class)

Entry	Size	Notes
deepseek-v4-flash-q2-q4	~91 GB	mixed q2/q4 (last 6 layers Q4_K), fits RAM, higher quality
deepseek-v4-flash-q4-ssd	~153 GB	full 4-bit experts, runs on 128 GB via SSD streaming
deepseek-v4-flash-q2-mtp	~81 GB	IQ2XXS + MTP speculative draft weights (greedy)
deepseek-v4-pro-q2-ssd	~433 GB	Pro-class, experimental SSD streaming

sha256s come from the HF LFS oids. yamllint passes.

Caveat

SSD streaming is Metal (Darwin) only upstream; the options are inert on CUDA/CPU. Documented in the -ssd gallery entries and .agents/ds4-backend.md.

Testing

The new wiring was type- and behavior-checked (-fsyntax-only + run harnesses: bare-flag, relative-path resolution, value validation, absolute passthrough) against the real ds4 headers at the pinned commit. A full cpu-ds4 link build was not run locally; CI backend-jobs is the first full compile.

Assisted-by: Claude:claude-opus-4-8 [Claude Code]