Update server.py

server.py

@@ -1,12 +1,5 @@
 """FastAPI backend for the LFM2.5 Spellchecker demo (Docker Space).
 
-Loads the published model from the Hub (pinned to `main`, so the Space always serves the current best
-checkpoint), exposes POST /api/correct, and serves the static frontend in static/. No Gradio.
-
-The model repo is private, so HF_TOKEN (a Space secret) is needed to pull it. Pinned library versions
-(see requirements.txt) match the environment the model was validated against — the encoder's custom
-bidirectional-mask code is sensitive to the transformers version.
-
     uvicorn server:app --host 0.0.0.0 --port 7860
 """
 import difflib
@@ -50,17 +43,11 @@ try:
 except RuntimeError:
     pass
 
-MODEL_ID = os.environ.get("SPELLCHECKER_MODEL", "LiquidAI/LFM2.5-Spellchecker-350M")
-# Pin to the EXACT published commit so the container can never serve stale cached weights/remote-code
-# (the bug we hit: a rebuild kept serving old, tagger-only behaviour). Bump on each publish, or override.
+MODEL_ID = os.environ.get("SPELLCHECKER_MODEL", "LiquidAI/LFM2.5-Encoder-350M-Spellchecker")
 MODEL_REV = os.environ.get("SPELLCHECKER_REVISION", "65a4a90af31205d2f7ef66b6a68d7b3d276adfdd")
 STATIC = os.path.join(os.path.dirname(os.path.abspath(__file__)), "static")
 
 print(f"[server] loading {MODEL_ID}@{MODEL_REV} on {_CPUS} CPU thread(s) ...", flush=True)
-# fp32, NOT fp16: x86 CPUs have no fast half-precision path, so .half() ran ~4.5x SLOWER here (matmuls
-# emulated per-op) for identical corrections. Uniform .float() casts every submodule — including the
-# reranker, which matches the tagger's dtype — so oneDNN/MKL use real f32 GEMM. int8 dynamic quant was
-# rejected: only ~15% faster than fp32 but it corrupts edits (GEC tagging is precision-sensitive).
 _model = AutoModel.from_pretrained(MODEL_ID, revision=MODEL_REV, trust_remote_code=True,
                                    token=os.environ.get("HF_TOKEN")).float().eval()
 _mem_bytes = sum(t.numel() * t.element_size() for t in (*_model.parameters(), *_model.buffers()))
@@ -100,9 +87,6 @@ def detok(text: str) -> str:
     return re.sub(r"\s+", " ", text).strip()
 
 
-# Startup self-test over the REAL user path (tokenize -> correct -> detok), logged + exposed at
-# /api/health: a correctly-deployed full system leaves this clean sentence UNCHANGED. If it changes,
-# the deploy is wrong (stale model, reranker inactive, or contraction handling broken) — no guessing.
 _PROBE_IN = "That's a fair point, let's discuss it tomorrow."
 try:
     _PROBE_OUT = detok(_model.correct([tokenize(_PROBE_IN)], max_iter=3)[0])