[ROCm][CI] Cleaning and restructuring amd-ci legacy pipeline (#34839)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-03-19 14:30:58 -05:00
committed by GitHub
parent 9279c59a0e
commit 040a505ff5
15 changed files with 2484 additions and 3552 deletions
+23 -1
View File
@@ -16,6 +16,23 @@ RAY_BASE_URL="https://raw.githubusercontent.com/ray-project/ray/master/python"
WORK_DIR=$(mktemp -d)
trap 'rm -rf "$WORK_DIR"' EXIT
# ── Detect PyTorch index URL ─────────────────────────────────────────────
if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
ROCM_VER=$(python3 -c "import torch; print(torch.version.hip.rsplit('.', 1)[0])")
CANDIDATE_URL="https://download.pytorch.org/whl/rocm${ROCM_VER}"
if curl -fsSL --head "${CANDIDATE_URL}/" >/dev/null 2>&1; then
TORCH_INDEX_URL="${CANDIDATE_URL}"
else
echo ">>> WARNING: ROCm ${ROCM_VER} wheel index not found at ${CANDIDATE_URL}"
echo ">>> Falling back to default PyPI (resolution may be incomplete)"
TORCH_INDEX_URL=""
fi
else
TORCH_INDEX_URL="https://download.pytorch.org/whl/cu129"
fi
echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
# Fetch all Ray requirement files used in the LLM depset pipeline
echo ">>> Fetching Ray requirement files"
RAY_FILES=(
@@ -116,6 +133,11 @@ echo "============================================================"
echo ">>> Resolving: Can Ray generate compatible lock files?"
echo "============================================================"
EXTRA_INDEX_ARGS=()
if [[ -n "${TORCH_INDEX_URL}" ]]; then
EXTRA_INDEX_ARGS+=(--extra-index-url "${TORCH_INDEX_URL}")
fi
set +e
uv pip compile \
"${WORK_DIR}/requirements.txt" \
@@ -126,7 +148,7 @@ uv pip compile \
-c "${WORK_DIR}/vllm-constraints.txt" \
--python-version 3.12 \
--python-platform x86_64-manylinux_2_31 \
--extra-index-url https://download.pytorch.org/whl/cu129 \
"${EXTRA_INDEX_ARGS[@]}" \
--index-strategy unsafe-best-match \
--unsafe-package setuptools \
--unsafe-package ray \
@@ -1,11 +1,14 @@
#!/usr/bin/env bash
set -euxo pipefail
# Nightly e2e test for prefetch offloading with a MoE model.
# Runs DeepSeek-V2-Lite with prefetch offloading of MoE expert weights
# and validates GSM8K accuracy matches baseline (no offloading).
#
# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
#
# Environment variables:
# ATTENTION_BACKEND - attention backend to use (e.g., FLASH_ATTN,
# ROCM_ATTN, FLASHINFER). If unset, uses vllm default.
THRESHOLD=${1:-0.25}
NUM_Q=${2:-1319}
PORT=${3:-8030}
@@ -22,6 +25,14 @@ wait_for_server() {
MODEL="deepseek-ai/DeepSeek-V2-Lite"
# ── Build optional vllm serve flags ─────────────────────────────────────
EXTRA_ARGS=()
if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
echo "Using attention backend: ${ATTENTION_BACKEND}"
EXTRA_ARGS+=(--attention-backend "${ATTENTION_BACKEND}")
fi
cleanup() {
if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
kill "${SERVER_PID}" 2>/dev/null || true
@@ -40,7 +51,8 @@ vllm serve "$MODEL" \
--offload-num-in-group 2 \
--offload-prefetch-step 1 \
--offload-params w13_weight w2_weight \
--port "$PORT" &
--port "$PORT" \
${EXTRA_ARGS+"${EXTRA_ARGS[@]}"} &
SERVER_PID=$!
wait_for_server "$PORT"