Sync from GitHub: dcdbd755536453bb053dd732d7d25cb3ad5a51c2

Dockerfile

@@ -41,6 +41,9 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
         build-essential \
         ca-certificates \
         curl \
+        pkg-config \
+        libcurl4-openssl-dev \
+        libssl-dev \
     && rm -rf /var/lib/apt/lists/*
 
 # HF Spaces convention: non-root user with UID 1000
@@ -57,27 +60,42 @@ ENV HOME=/home/user \
 WORKDIR /home/user/bitnet
 RUN git clone --recursive https://github.com/microsoft/BitNet.git /home/user/bitnet
 
-# Install BitNet's Python build/utility deps
+# Install BitNet's Python build/utility deps (gguf, huggingface_hub,
+# numpy, torch-cpu, etc. — used by their conversion + download scripts,
+# which we're about to bypass for the compile, but keep for ggml tools).
 RUN pip install --no-cache-dir --user -r /home/user/bitnet/requirements.txt
 
-# Build + download BitNet 2B4T (i2_s = ternary quant scheme specific to
-# BitNet's kernels). This single call: downloads the GGUF weights,
-# preprocesses them, configures cmake, and compiles llama-cli.
+# Compile llama-cli directly via cmake. Bypasses setup_env.py (which
+# swallows cmake errors into logs/compile.log with no way to see them
+# on build failure). -j 2 limits parallel compile jobs so LLVM kernel
+# compilation doesn't OOM the build runner. On compile failure, dump
+# all cmake + compile logs so next-iteration debugging has signal.
 RUN cd /home/user/bitnet \
-    && python setup_env.py --hf-repo microsoft/BitNet-b1.58-2B-4T -q i2_s
+    && cmake -B build \
+         -DCMAKE_BUILD_TYPE=Release \
+         -DBITNET_X86_TL2=ON \
+    && (cmake --build build -j 2 --target llama-cli || \
+        (echo "==================== BUILD FAILED ====================" && \
+         find /home/user/bitnet -name "CMakeError.log" -o -name "CMakeOutput.log" -o -name "compile.log" 2>/dev/null | \
+           while read f; do echo "=== $f ===" && cat "$f" || true; done ; \
+         exit 1))
 
-# Download Falcon3-10B-Instruct 1.58bit GGUF (already ternary-quantized
-# by tiiuae; we just need the weights). Filename inside the repo gets
-# resolved at runtime by globbing for "*.gguf" in the dir.
+# Download GGUF weights directly via huggingface_hub (decoupled from
+# the BitNet repo's preprocess scripts since those assume a specific
+# directory layout). Both models' GGUF files land in known dirs that
+# BitnetCppClient.resolve_gguf finds at runtime.
 RUN python -c "\
 from huggingface_hub import snapshot_download; \
+snapshot_download('microsoft/bitnet-b1.58-2B-4T-gguf', \
+    local_dir='/home/user/models/bitnet-2b-gguf', \
+    allow_patterns=['*.gguf']); \
 snapshot_download('tiiuae/Falcon3-10B-Instruct-1.58bit-GGUF', \
     local_dir='/home/user/models/falcon3-10b-gguf', \
     allow_patterns=['*.gguf'])"
 
 # Paths exposed to app.py via env vars
 ENV BITNET_CPP_BINARY=/home/user/bitnet/build/bin/llama-cli
-ENV BITNET_CHAT_GGUF_DIR=/home/user/bitnet/models
+ENV BITNET_CHAT_GGUF_DIR=/home/user/models/bitnet-2b-gguf
 ENV FALCON_EXTRACTOR_GGUF_DIR=/home/user/models/falcon3-10b-gguf
 
 # ── Python app deps + repo ──────────────────────────────────────────