fix: remove kvmanager feature from python 3.12 ai-dynamo-runtime wheel (#2456)

Cargo.lock

@@ -1814,7 +1814,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-engine-llamacpp"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "async-stream",
  "dynamo-llm",
@@ -1826,7 +1826,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-engine-mistralrs"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "anyhow",
  "async-openai",
@@ -1844,7 +1844,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-llm"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "ahash",
  "akin",
@@ -1923,7 +1923,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-run"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "anyhow",
  "async-openai",
@@ -1952,7 +1952,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-runtime"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "anyhow",
  "arc-swap",
@@ -2009,7 +2009,7 @@ dependencies = [
 
 [[package]]
 name = "dynamo-tokens"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "bytemuck",
  "derive-getters",
@@ -3776,7 +3776,7 @@ checksum = "d750af042f7ef4f724306de029d18836c26c1765a54a6a3f094cbd23a7267ffa"
 
 [[package]]
 name = "libdynamo_llm"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "anyhow",
  "async-once-cell",
@@ -4088,7 +4088,7 @@ dependencies = [
 
 [[package]]
 name = "metrics"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "axum 0.8.3",
  "clap 4.5.40",
@@ -5899,7 +5899,7 @@ dependencies = [
 
 [[package]]
 name = "router"
-version = "0.4.0"
+version = "0.4.0+post0"
 dependencies = [
  "clap 4.5.40",
  "dynamo-llm",

Cargo.toml

@@ -15,7 +15,7 @@ members = [
 resolver = "3"
 
 [workspace.package]
-version = "0.4.0"
+version = "0.4.0+post0"
 edition = "2021"
 description = "Dynamo Inference Framework"
 authors = ["NVIDIA Inc. <sw-dl-dynamo@nvidia.com>"]

container/Dockerfile

@@ -8,6 +8,7 @@ ARG BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
 # can be updated to later versions.
 ARG BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
 ARG RELEASE_BUILD=false
+ARG ENABLE_KVBM=false
 
 # Define general architecture ARGs for supporting both x86 and aarch64 builds.
 #   ARCH: Used for package suffixes (e.g., amd64, arm64)
@@ -197,6 +198,8 @@ ARG CARGO_BUILD_JOBS
 ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16}
 # Use build arg RELEASE_BUILD = true to generate wheels for Python 3.10, 3.11 and 3.12.
 ARG RELEASE_BUILD
+# Use arg ENABLE_KVBM = true to turn on the block-manager feature
+ARG ENABLE_KVBM
 
 WORKDIR /opt/dynamo
 
@@ -228,7 +231,11 @@ COPY components/ /opt/dynamo/components/
 RUN uv build --wheel --out-dir /opt/dynamo/dist && \
     cd /opt/dynamo/lib/bindings/python && \
     uv pip install maturin[patchelf] && \
-    maturin build --release --features block-manager --out /opt/dynamo/dist && \
+    if [ "$ENABLE_KVBM" = "true" ]; then \
+        maturin build --release --features block-manager --out /opt/dynamo/dist; \
+    else \
+        maturin build --release --out /opt/dynamo/dist; \
+    fi && \
     if [ "$RELEASE_BUILD" = "true" ]; then \
         # do not enable KVBM feature, ensure compatibility with lower glibc
         uv run --python 3.11 maturin build --release --out /opt/dynamo/dist && \

container/Dockerfile.vllm

@@ -8,6 +8,7 @@ ARG BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
 # can be updated to later versions.
 ARG BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
 ARG RELEASE_BUILD
+ARG ENABLE_KVBM=false
 ARG RUNTIME_IMAGE="nvcr.io/nvidia/cuda"
 ARG RUNTIME_IMAGE_TAG="12.8.1-runtime-ubuntu24.04"
 
@@ -320,6 +321,8 @@ ARG CARGO_BUILD_JOBS
 ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16}
 # Use build arg RELEASE_BUILD = true to generate wheels for Python 3.10, 3.11 and 3.12.
 ARG RELEASE_BUILD
+# Use arg ENABLE_KVBM = true to turn on the block-manager feature
+ARG ENABLE_KVBM
 
 # Keep in sync with the base image.
 ENV NIXL_PREFIX=/opt/nvidia/nvda_nixl
@@ -369,7 +372,11 @@ RUN cargo build \
 RUN uv build --wheel --out-dir /workspace/dist && \
     cd /workspace/lib/bindings/python && \
     uv pip install maturin[patchelf] && \
-    maturin build --release --features block-manager --out /workspace/dist && \
+    if [ "$ENABLE_KVBM" = "true" ]; then \
+        maturin build --release --features block-manager --out /workspace/dist; \
+    else \
+        maturin build --release --out /workspace/dist; \
+    fi && \
     if [ "$RELEASE_BUILD" = "true" ]; then \
         # do not enable KVBM feature, ensure compatibility with lower glibc
         uv run --python 3.11 maturin build --release --out /workspace/dist && \

container/build.sh

@@ -274,6 +274,9 @@ get_options() {
         --release-build)
             RELEASE_BUILD=true
             ;;
+        --enable-kvbm)
+            ENABLE_KVBM=true
+            ;;
         --make-efa)
             NIXL_UCX_REF=$NIXL_UCX_EFA_REF
             ;;
@@ -530,6 +533,11 @@ if [  ! -z ${RELEASE_BUILD} ]; then
     BUILD_ARGS+=" --build-arg RELEASE_BUILD=${RELEASE_BUILD} "
 fi
 
+if [  ! -z ${ENABLE_KVBM} ]; then
+    echo "Enabling the KVBM in the ai-dynamo-runtime"
+    BUILD_ARGS+=" --build-arg ENABLE_KVBM=${ENABLE_KVBM} "
+fi
+
 if [ -n "${NIXL_UCX_REF}" ]; then
     BUILD_ARGS+=" --build-arg NIXL_UCX_REF=${NIXL_UCX_REF} "
 fi

lib/bindings/python/Cargo.toml

@@ -19,7 +19,7 @@
 
 [package]
 name = "dynamo-py3"
-version = "0.4.0"
+version = "0.4.0+post0"
 edition = "2021"
 authors = ["NVIDIA"]
 license = "Apache-2.0"

lib/bindings/python/pyproject.toml

@@ -16,7 +16,7 @@
 
 [project]
 name = "ai-dynamo-runtime"
-dynamic = ["version"]
+version = "0.4.0.post0"
 description = "Dynamo Inference Framework Runtime"
 readme = "README.md"
 authors = [

lib/bindings/python/tests/test_block_manager.py

@@ -19,7 +19,13 @@ import asyncio
 import pytest
 import torch
 
-from dynamo.llm import BlockManager
+# Attempt to import the optional module
+try:
+    from dynamo.llm import BlockManager
+except ImportError:
+    pytest.importorskip(
+        "optional_module", reason="block-manager feature is not enabled"
+    )
 
 pytestmark = pytest.mark.pre_merge
 

lib/runtime/examples/Cargo.toml

@@ -22,7 +22,7 @@ members = [
 resolver = "3"
 
 [workspace.package]
-version = "0.4.0"
+version = "0.4.0+post0"
 edition = "2021"
 authors = ["NVIDIA"]
 license = "Apache-2.0"

pyproject.toml

@@ -3,7 +3,7 @@
 
 [project]
 name = "ai-dynamo"
-version = "0.4.0"
+version = "0.4.0.post0"
 description = "Distributed Inference Framework"
 readme = "README.md"
 authors = [
@@ -13,7 +13,7 @@ license = { text = "Apache-2.0" }
 license-files = ["LICENSE"]
 requires-python = ">=3.10"
 dependencies = [
-    "ai-dynamo-runtime==0.4.0",
+    "ai-dynamo-runtime==0.4.0.post0",
     "pytest>=8.3.4",
     "types-psutil>=7.0.0.20250218",
     "kubernetes>=32.0.1,<33.0.0",