fix: b12x non-editable install (files land in site-packages)

docker-images/glm52-nf3-hybrid/Dockerfile

@@ -100,7 +100,7 @@ ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST} CUTE_DSL_ARCH=sm_121a
 RUN git clone --recursive ${B12X_REPO} /opt/b12x \
  && cd /opt/b12x \
  && git checkout ${B12X_REF} \
- && pip install --no-cache-dir -e . --no-build-isolation \
+ && pip install --no-cache-dir . --no-build-isolation \
     -C setup.args="--cuda-arch=${TORCH_CUDA_ARCH_LIST}"
 
 FROM base AS runner

recipes/glm52-nf3-hybrid-context/recipe.toml

@@ -48,13 +48,14 @@ distributed_backend = "mp"
 shm_size = "64gb"
 docker_image_dir = "glm52-nf3-hybrid"
 
-# entrypoint = ["vllm", "serve"]
+entrypoint = ["vllm", "serve"]
 
 args = [
     # Quantization & KV cache
     "--quantization", "modelopt_fp4",
     "--kv-cache-dtype", "fp8",
-    "--load-format", "fastsafetensors",
+    "--load-format", "auto",
+#    "--load-format", "fastsafetensors",
     # Context length (800K max-len, 877K KV pool)
     "--max-model-len", "800000",
     # Batch & memory