fix: switch to voipmonitor/b12x@f3686b55 (eldritch stack)

docker-images/glm52-nf3-hybrid/Dockerfile

@@ -97,7 +97,11 @@ FROM base AS b12x-builder
 ARG B12X_REPO B12X_REF B12X_COMMIT
 ARG TORCH_CUDA_ARCH_LIST=12.1a
 ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST} CUTE_DSL_ARCH=sm_121a
-RUN pip install --no-cache-dir "b12x @ git+${B12X_REPO}@${B12X_REF}"
+RUN git clone --recursive ${B12X_REPO} /opt/b12x \
+ && cd /opt/b12x \
+ && git checkout ${B12X_REF} \
+ && pip install --no-cache-dir -e . --no-build-isolation \
+    -C setup.args="--cuda-arch=${TORCH_CUDA_ARCH_LIST}"
 
 FROM base AS runner
 ARG PYTORCH_INDEX_URL PYPI_INDEX_URL

docker-images/glm52-nf3-hybrid/versions.env

@@ -33,9 +33,9 @@ VLLM_REF=codex/dcp-globaltopk-sharddraft-defaults-20260622
 VLLM_COMMIT=e232d262369b8c918cf478a7a96a0fcf8127cf65
 
 # b12x (NVFP4 kernel library)
-B12X_REPO=https://github.com/lukealonso/b12x.git
-B12X_REF=9cd63a7
-B12X_COMMIT=9cd63a7dccd877df41f9191ec6a3aa92c03db652
+B12X_REPO=https://github.com/voipmonitor/b12x.git
+B12X_REF=f3686b55
+B12X_COMMIT=f3686b555d639823b276c2080f173145eed7f007
 
 # FlashInfer
 FLASHINFER_REPO=https://github.com/flashinfer-ai/flashinfer.git

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

@@ -48,7 +48,7 @@ distributed_backend = "mp"
 shm_size = "64gb"
 docker_image_dir = "glm52-nf3-hybrid"
 
-entrypoint = ["vllm", "serve"]
+# entrypoint = ["vllm", "serve"]
 
 args = [
     # Quantization & KV cache
@@ -61,7 +61,7 @@ args = [
     "--max-num-seqs", "4",
     "--max-num-batched-tokens", "4096",
     "--max-cudagraph-capture-size", "32",
-    "--gpu-memory-utilization", "0.88",
+    "--gpu-memory-utilization", "0.92",
     # Engine backends
     "--attention-backend", "B12X_MLA_SPARSE",
     "--moe-backend", "b12x",