Add CodeBuild CI + minimal SGLang overlay

ci/buildspec.yml

@@ -0,0 +1,56 @@
+# AWS CodeBuild buildspec for megatron-deepep-v2-efa.
+# Mirroring vllm-deepep-v2-efa pattern (Wave 5).
+version: 0.2
+
+env:
+  variables:
+    DOCKER_BUILDKIT: "1"
+    BASE_IMAGE_FAST: "ghcr.io/antonai-work/deepep-v2-efa-base:v0.2.5-sm90a"
+
+phases:
+  install:
+    commands:
+      - echo "[install] Installing helpers"
+      - docker --version && aws --version
+
+  pre_build:
+    commands:
+      - echo "[pre_build] Starting $(date -u +%Y-%m-%dT%H:%M:%SZ)"
+      - test -n "${AWS_ACCOUNT_ID}" || (echo "ERROR: AWS_ACCOUNT_ID unset" && exit 2)
+      - test -n "${AWS_REGION}" || (echo "ERROR: AWS_REGION unset" && exit 2)
+      - test -n "${ECR_REPO}" || (echo "ERROR: ECR_REPO unset" && exit 2)
+      - ECR_REGISTRY="${AWS_ACCOUNT_ID}.dkr.ecr.${AWS_REGION}.amazonaws.com"
+      - aws ecr get-login-password --region "${AWS_REGION}" | docker login --username AWS --password-stdin "${ECR_REGISTRY}"
+      - |
+        if [ -n "${GHCR_READ_TOKEN_SECRET_ARN:-}" ]; then
+          echo "[pre_build] Logging in to GHCR"
+          GHCR_TOKEN="$(aws secretsmanager get-secret-value --secret-id "${GHCR_READ_TOKEN_SECRET_ARN}" --query SecretString --output text)"
+          echo "${GHCR_TOKEN}" | docker login ghcr.io --username antonai-work --password-stdin
+          unset GHCR_TOKEN
+        fi
+      - GIT_SHA_SHORT="$(git rev-parse --short=12 HEAD)"
+      - test -f pins.env && . ./pins.env || true
+      - docker pull "${BASE_IMAGE_FAST}"
+
+  build:
+    commands:
+      - FAST_TAG_SHA="${ECR_REGISTRY}/${ECR_REPO}:fast-${GIT_SHA_SHORT}"
+      - FAST_TAG_LATEST="${ECR_REGISTRY}/${ECR_REPO}:fast-latest"
+      - echo "[build] Building ${FAST_TAG_SHA}"
+      - |
+        docker build -f docker/Dockerfile \
+          --build-arg BUILD_MODE=fast \
+          --build-arg BASE_IMAGE_FAST="${BASE_IMAGE_FAST}" \
+          -t "${FAST_TAG_SHA}" -t "${FAST_TAG_LATEST}" .
+
+  post_build:
+    commands:
+      - echo "[post_build] Starting $(date -u +%Y-%m-%dT%H:%M:%SZ)"
+      - |
+        if [ "${CODEBUILD_BUILD_SUCCEEDING:-0}" != "1" ]; then
+          echo "[post_build] Build failing; skipping push"
+          exit 0
+        fi
+      - docker push "${FAST_TAG_SHA}"
+      - docker push "${FAST_TAG_LATEST}"
+      - echo "[post_build] Done"

docker/Dockerfile

@@ -42,9 +42,17 @@ SHELL ["/bin/bash", "-o", "pipefail", "-c"]
 # -----------------------------------------------------------------------------
 ARG SGLANG_VERSION=0.5.6.post2
 
-# TODO (Wave 29): install sglang + any SGLang-specific deep_ep adapter and
-# re-pin torch/nccl/nvshmem to match the v0.2.5 base (see vllm-deepep-v2-efa
-# Dockerfile for the reference pin-guard sequence).
+# Minimal overlay for Wave 29 cross-node smoke. Full DeepEP V2 adapter
+# (sgl-project/sglang#24443) + DeepGEMM JIT cache baking TBD.
+RUN pip install --no-cache-dir --break-system-packages "sglang[all]==${SGLANG_VERSION}"
+
+# Guard torch/nccl/nvshmem to match v0.2.5 base (cu13-unified stack).
+# Any torch upgrade silently re-installs nvidia-nccl/nvshmem wheels.
+ARG NVIDIA_NCCL_PIN=2.30.4
+ARG NVSHMEM_PIN=2.11.2
+RUN pip install --no-cache-dir --break-system-packages --no-deps --force-reinstall \
+      "nvidia-nccl-cu13==${NVIDIA_NCCL_PIN}" \
+      "nvidia-nvshmem-cu13==${NVSHMEM_PIN}"
 
 ENV PYTHONPATH=/opt
 

pins.env

@@ -1,20 +1,7 @@
-# Canonical pins for sglang-deepep-v2-efa.
-# Last updated: 2026-05-07 (scaffolded from Wave 28 cascade; Wave 29 pending).
-
-BASE_IMAGE_FAST=ghcr.io/antonai-work/deepep-v2-efa-base:v0.2.5-sm90a
-
-# DeepEP V2 (already baked into BASE_IMAGE_FAST; listed here for audit only)
-DEEPEP_FORK=https://github.com/dmvevents/DeepEP-1.git
-DEEPEP_BRANCH=aws-efa-auto-qp-cap-v2
-DEEPEP_SHA=146cc356aa00c39ac1590c05775e05b0f031e70c
-
-# SGLang target release (matches V1 sglang-deepep-working:v3 source line in
-# awesome-inferencing). Bump after validating native ElasticBuffer (Wave 29).
-SGLANG_VERSION=0.5.6.post2
-
-# Upstream SGLang DeepEP V2 PR — re-validate on v0.2.5 in Wave 29.
-SGLANG_UPSTREAM_PR=24443
-SGLANG_UPSTREAM_HEAD_SHA=f66ab63b499b7671ee8d7052b2e54708a219ddb1
-
-# Wave 13 cu13 cascade alignment (see parent base for rationale).
-NVIDIA_NCCL_PIN="nvidia-nccl-cu13>=2.30.4"
+# Canonical pin source for sglang-deepep-v2-efa (Wave 29).
+export BASE_IMAGE_FAST="ghcr.io/antonai-work/deepep-v2-efa-base:v0.2.5-sm90a"
+export DEEPEP_FORK="https://github.com/deepseek-ai/DeepEP.git"
+export DEEPEP_BRANCH="aws-efa-auto-qp-cap-v2"
+export DEEPEP_SHA="146cc356aa00c39ac1590c05775e05b0f031e70c"
+export NVIDIA_NCCL_PIN="2.30.4"
+export NVSHMEM_PIN="2.11.2"