fix(runtime): finalize HF model download (create snapshot symlink) — fixes model-not-found on every binary

[SnowCheetos]

Critical: HF model download is broken in every shipped binary

Any run that downloads a model from Hugging Face fails at warmup:

runtime warmup failed: embedding error: embedding model not found after resolution:
.../models--ggml-org--Qwen3-1.7B-GGUF/snapshots/<rev>/Qwen3-1.7B-Q4_K_M.gguf

The download succeeds — the full GGUF blob lands in blobs/<sha> — but the resolved path (snapshots/<rev>/<file>) is a symlink that never gets created, so ModelHandle::load's path.exists() check fails.

Root cause

resolve_hf_model in the bridge does:

auto plan = common_download_get_hf_plan(model, opts);
common_download_run_tasks({ common_download_task(plan.primary, opts) });   // writes blobs/<sha>
model.path = plan.primary.final_path… ;                                    // ← dangling snapshot path

Since b9837, writing the blob and creating the snapshots/<rev>/<file> → ../../blobs/<sha> symlink are separate steps. The symlink + path selection is done by the public hf_cache::finalize_file() (hf-cache.cpp). llama.cpp's own common/arg.cpp does get_hf_plan → run_tasks → finalize_file; the bridge skipped that last step.

Fix

Call hf_cache::finalize_file(plan.primary) after the download and use its return value — it creates the snapshot symlink (or returns the blob path directly if symlinks are unavailable) and hands back the path to load. Exactly mirrors upstream arg.cpp.

+#include "hf-cache.h"
 ...
-model.path = plan.primary.final_path.empty() ? plan.primary.local_path : plan.primary.final_path;
+model.path = hf_cache::finalize_file(plan.primary);

Verification

• Bridge compiles against a real b9837 checkout in both prebuilt and source modes; finalize_file resolves from hf-cache.h.
• Confirmed the failure mode on-box: the blob was a valid GGUF (GGUF magic) sitting in blobs/ with an empty snapshots/; manually creating the exact symlink finalize_file would make lets the resolved path load.

Impact / merge priority

This affects every backend, including the released prebuilt binaries — HF --hf-repo downloads (the default onboarding path via autocommit init) are broken without it. Worth merging ahead of a v0.17.6 tag.

[Copilot]

Pull request overview

Fixes Hugging Face model downloads in the llama.cpp bridge by ensuring the resolved snapshot path is actually materialized after downloading (preventing “model not found after resolution” failures in warmup for shipped binaries).

Changes:

• Include llama.cpp’s HF cache API header (hf-cache.h) in the bridge.
• After common_download_run_tasks, call hf_cache::finalize_file(plan.primary) and use its returned path as the load path (creating the snapshots/<rev>/<file> -> ../../blobs/<sha> symlink when supported).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.