betterwithage commited on
Commit
2b1c616
·
verified ·
1 Parent(s): 114fef3

chore(sync): mirror backend .py + Dockerfile to Space (hf-sync-backend)

Browse files

Automated backend sync from szl-holdings/a11oy main via hf-sync-backend.
Updated (differed from the Space): Dockerfile

Keeps the Space-built backend (serve.py + the Dockerfile-COPY'd .py
modules) identical to GitHub main so the Space never rebuilds from a
stale backend and new endpoints don't 404 there.

Files changed (1) hide show
  1. Dockerfile +6 -0
Dockerfile CHANGED
@@ -429,6 +429,12 @@ COPY szl_alloy_models.py ./szl_alloy_models.py
429
  # official CPU wheel index, so this resolves deterministically with NO from-source
430
  # build. The `|| echo` stays ONLY as a final safety net for a platform without a
431
  # matching prebuilt wheel (honest tower-side fallback, never fabricated output).
 
 
 
 
 
 
432
  RUN pip install --no-cache-dir \
433
  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu \
434
  "llama-cpp-python==0.3.19" \
 
429
  # official CPU wheel index, so this resolves deterministically with NO from-source
430
  # build. The `|| echo` stays ONLY as a final safety net for a platform without a
431
  # matching prebuilt wheel (honest tower-side fallback, never fabricated output).
432
+ # GUARD: because the `|| echo` keeps the image build GREEN even if this wheel
433
+ # vanishes for cp312/linux_x86_64 (a silent demo-tier degrade), CI workflow
434
+ # .github/workflows/llama-wheel-guard.yml re-runs this exact pinned install
435
+ # (parsed from this line) on cp312/linux_x86_64 WITHOUT the mask and with
436
+ # --only-binary, failing loudly if the prebuilt wheel is gone. When bumping the
437
+ # version below, that guard verifies — does not assume — the new wheel exists.
438
  RUN pip install --no-cache-dir \
439
  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu \
440
  "llama-cpp-python==0.3.19" \