From 1327560ea72e928d01e9b10f98ed233ac30c0b1a Mon Sep 17 00:00:00 2001
From: thxCode <thxcode0824@gmail.com>
Date: Tue, 1 Sep 2026 16:48:01 +0800
Subject: [PATCH] feat(dtk): post-operation to pin numpy and remove nixl ep for
 vllm 0.18.1

- add pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep with matrix
  rule and overlay Dockerfile porting PR #265 fixes
- log the operation in pack/.post_operation/README.md
- ignore local .claude/ tooling directory

Task 1 of dtk-post-op-pin-numpy-remove-nixl.

Signed-off-by: thxCode <thxcode0824@gmail.com>
---
 .gitignore                                    |  3 +
 .../dtk/Dockerfile                            | 67 +++++++++++++++++++
 .../matrix.yaml                               | 16 +++++
 pack/.post_operation/README.md                |  1 +
 4 files changed, 87 insertions(+)
 create mode 100644 pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/dtk/Dockerfile
 create mode 100644 pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/matrix.yaml

diff --git a/.gitignore b/.gitignore
index e96aa10a..2eb85dc5 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,6 +4,9 @@
 .DS_Store
 *.swp
 
+# Local agent tooling
+.claude/
+
 # C extensions
 *.so
 
diff --git a/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/dtk/Dockerfile b/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/dtk/Dockerfile
new file mode 100644
index 00000000..e0210e5c
--- /dev/null
+++ b/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/dtk/Dockerfile
@@ -0,0 +1,67 @@
+ARG DTK_VERSION=26.04
+ARG VLLM_VERSION=0.18.1
+
+FROM gpustack/runner:dtk${DTK_VERSION}-vllm${VLLM_VERSION} AS vllm
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Restore NumPy 1.x ABI compatibility
+
+# The DTK base image currently ships a Torch build compiled against the NumPy
+# 1.x ABI. NumPy 2.x cannot initialize in that Torch build, which causes
+# distributed initialization to fail when it calls Tensor.numpy().
+#
+# Install without dependencies to avoid changing the prebuilt DTK Python
+# environment beyond the NumPy compatibility pin.
+RUN <<EOF
+    uv pip install --no-deps "numpy==1.26.4"
+
+    python - <<'PYEOF'
+import numpy
+
+assert numpy.__version__ == "1.26.4", \
+    f"expected numpy 1.26.4, got {numpy.__version__}"
+PYEOF
+EOF
+
+## Remove CUDA-only NIXL EP packages
+
+# The prebuilt DTK vLLM base image can inherit NIXL Expert Parallel packages
+# from its Python environment. NIXL EP provides NVIDIA CUDA backends which
+# require libcuda.so.1 and cannot run on Hygon DCU systems.
+#
+# vLLM detects NIXL EP through importlib.util.find_spec("nixl_ep"). Remove
+# the distributions entirely, rather than leaving an ImportError stub, so
+# vLLM selects its non-NIXL path. Use uv to resolve the active Python
+# environment and package paths instead of relying on a fixed site-packages
+# location.
+RUN <<EOF
+    # The NIXL EP modules are owned by the nixl, nixl-cu12, and nixl-cu13
+    # distributions in the DTK base image. Do not fail if a future base image
+    # no longer includes one of them; the import check below remains required.
+    uv pip uninstall nixl nixl-cu12 nixl-cu13 || true
+
+    python - <<'PYEOF'
+import importlib.util
+
+for module_name in ("nixl_ep", "nixl_ep_cu12", "nixl_ep_cu13"):
+    assert importlib.util.find_spec(module_name) is None, \
+        f"{module_name} is still discoverable after DTK image cleanup"
+PYEOF
+EOF
+
+RUN <<EOF
+    # Review
+    uv pip tree \
+        --package numpy \
+        --package torch \
+        --package vllm
+EOF
+
+## Entrypoint
+
+WORKDIR /
+ENTRYPOINT [ "tini", "--" ]
diff --git a/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/matrix.yaml b/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/matrix.yaml
new file mode 100644
index 00000000..093b566e
--- /dev/null
+++ b/pack/.post_operation/20260901_dtk_pin_numpy_remove_nixl_ep/matrix.yaml
@@ -0,0 +1,16 @@
+rules:
+
+  #
+  # Hygon DTK
+  #
+
+  ## Packed Hygon DTK 26.04.
+  ##
+  - backend: "dtk"
+    services:
+      - "vllm"
+    platforms:
+      - "linux/amd64"
+    args:
+      - "DTK_VERSION=26.04"
+      - "VLLM_VERSION=0.18.1"
diff --git a/pack/.post_operation/README.md b/pack/.post_operation/README.md
index f00e6105..e419c450 100644
--- a/pack/.post_operation/README.md
+++ b/pack/.post_operation/README.md
@@ -44,3 +44,4 @@ We leverage the matrix expansion feature of GPUStack Runner to achieve this, and
 - [ ] 2026-02-14: Patch SGLang 0.5.8.post1 of CANN/CUDA/ROCm released images to reduce Z-Image loading memory occupation.
 - [x] 2026-02-28: Reinstall `vllm-omni` packages for vLLM 0.16.0 of CUDA released images.
 - [x] 2026-03-03: Fix malformed ARM64 image for vLLM 0.15.1 of CUDA released images.
+- [ ] 2026-09-01: Pin `numpy` to 1.26.4 and remove CUDA-only NIXL EP packages for vLLM 0.18.1 of DTK 26.04 released images.