#!/usr/bin/env bash
# Gabriele's instruction, executed verbatim.
#
#   export FASTKERNELS_RESULTS_DIR=<path to a folder on your fsx>
#   fastkernels validate full
#
# No --output-dir, no --timeout/--stall-timeout overrides, no serial override:
# Ray schedules the jobs in parallel, which is what the instruction describes
# ("This will launch a Ray server with all the jobs in parallel").
#
# Differences from the earlier 20260922 run, all deliberate:
#   - runs out of ~/fastkernels, a pristine clone of main @ 9acebaa, so none of
#     the uncommitted local changes in /code/.../project/fastkernels apply
#     (no Whisper 30s clamp, no FASTKERNELS_VALIDATE_SERIAL switch)
#   - parallel instead of serial
#   - stock timeouts
set -uo pipefail

# What "log out and back in" accomplishes: the HF token and cache location that
# setup_node.sh wrote into ~/.bashrc. Set them here because ~/.bashrc returns
# early for non-interactive shells.
export HF_HOME=/home/yak/data-fast/huggingface
export HF_XET_HIGH_PERFORMANCE=1
ulimit -c 0
# vLLM tries to parse anything named VLLM_*; drop the K8s service vars.
unset $(env | cut -d= -f1 | grep -E '^(SOYOUNG_)?(VLLM|QWEN)_[A-Z0-9]+_.*(SERVICE|PORT)') 2>/dev/null

export FASTKERNELS_RESULTS_DIR=/code/users/alexsu/project/fastkernels/experiments/20260923-validate-full-h200-strict

cd ~/fastkernels
nvidia-smi -L || { echo "GPU not visible, aborting"; exit 1; }

set -x
fastkernels validate full