fix(sglang): ship nvtx package and nsys PATH for DYN_NVTX profiling (#8629)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

container/deps/requirements.common.txt

@@ -16,6 +16,8 @@ matplotlib==3.10.7
 msgpack==1.1.2
 msgspec==0.19.0
 nvidia-ml-py<=13.580.65  # NVIDIA/CUDA related, may vary by driver version
+# NVTX Python bindings used by dynamo.common.utils.nvtx_utils for DYN_NVTX=1 profiling.
+nvtx==0.2.14
 opentelemetry-api<=1.38.0  # Pinned for tracing support
 opentelemetry-exporter-otlp<=1.38.0
 opentelemetry-sdk<=1.38.0

container/deps/requirements.vllm.txt

@@ -5,7 +5,6 @@
 
 av==15.0.0
 ftfy==6.3.1
-nvtx==0.2.14
 # Pin transformers and onnxruntime: vLLM only sets a lower bound on
 # transformers, so resolver drift can change versions between builds
 # and shift LLaVA/vision preprocessing, making multimodal outputs

container/templates/sglang_runtime.Dockerfile

@@ -68,6 +68,15 @@ RUN --mount=type=cache,target=/root/.cache/pip,sharing=locked \
     fi
 {% endif %}
 
+# Install nvtx pinned in container/deps/requirements.common.txt so DYN_NVTX=1
+# profiling works in all targets (runtime, dev, local-dev) — see
+# components/src/dynamo/common/utils/nvtx_utils.py. --no-deps preserves the
+# upstream lmsysorg/sglang Python stack.
+RUN --mount=type=bind,source=./container/deps/requirements.common.txt,target=/tmp/requirements.common.txt \
+    --mount=type=cache,target=/root/.cache/pip,sharing=locked \
+    export PIP_CACHE_DIR=/root/.cache/pip && \
+    pip install --break-system-packages --no-deps $(grep -E '^nvtx==' /tmp/requirements.common.txt)
+
 # Copy tests, deploy and components for CI with correct ownership
 COPY --chmod=775 --chown=dynamo:0 tests /workspace/tests
 COPY --chmod=775 --chown=dynamo:0 examples /workspace/examples
@@ -87,7 +96,10 @@ RUN --mount=type=bind,source=./container/launch_message/runtime.txt,target=/opt/
 
 RUN chmod 755 /opt/dynamo/.launch_screen && \
     echo 'cat /opt/dynamo/.launch_screen' >> /etc/bash.bashrc && \
-    ln -s /workspace /sgl-workspace/dynamo
+    ln -s /workspace /sgl-workspace/dynamo && \
+    NSYS_BIN=$(find /opt/nvidia/nsight-compute -maxdepth 6 -type f -name nsys -executable 2>/dev/null | head -n1) && \
+    if [ -n "$NSYS_BIN" ]; then ln -sf "$NSYS_BIN" /usr/local/bin/nsys; \
+    else echo "WARNING: no bundled nsys found under /opt/nvidia/nsight-compute"; fi
 
 USER dynamo
 ARG DYNAMO_COMMIT_SHA