docker pull adaptfm/adaptfm-base:latest

pip install huggingface_hub
python3 -c "from huggingface_hub import snapshot_download; snapshot_download('Qwen/Qwen3.5-4B', local_dir='./qwen-weights')"

docker run -d --gpus all -p 8080:8080 --name test-run \
  -v hf_cache:/root/.cache/huggingface --shm-size=4g \
  adaptfm/adaptfm-base:latest

# Wait ~3 min for model to load, then test
curl http://localhost:8080/ping

# Test with thinking disabled (MMLU-Pro/IFEval mode)
curl -s -X POST http://localhost:8080/invocations \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"What is 2+2?"}],"max_tokens":32}' | python3 -m json.tool

# Test with thinking enabled (GPQA mode)
curl -s -X POST http://localhost:8080/invocations \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"What is 2+2?"}],"max_tokens":512,"thinking":true}' | python3 -m json.tool

docker rm -f test-run

docker build -t my-submission:latest .

# vLLM example
--chat-template /path/to/qwen_no_think.jinja

# Start your container
docker run -d --gpus all -p 8080:8080 --name test-submission my-submission:latest
watch -n5 'curl -s http://localhost:8080/ping'  # wait for 200

# Quality only (~20 min with 10% sample)
HF_HOME=/path/to/hf_cache QUALITY_LIMIT=0.1 NUM_CONCURRENT=8 \
  python3 run_quality_local.py 2>&1 | tee /tmp/quality.log

# Full eval — latency + quality (~60 min with 10% sample)
HF_HOME=/path/to/hf_cache EVAL_MODE=full QUALITY_LIMIT=0.1 NUM_CONCURRENT=8 \
  python3 run_eval_local.py 2>&1 | tee /tmp/eval.log

curl -s -X POST \
  https://79x0as8g44.execute-api.us-east-1.amazonaws.com/prod/register \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: qoXdZQNYbX1s4wnhmcBJG2APABlqjNVSao8CdM3j' \
  -d '{"team_id": "AFM-xxxxxxxx"}' | python3 -m json.tool

docker save my-submission:latest | gzip > image.tar.gz
du -sh image.tar.gz  # must be under 20 GB

FILE_SIZE=$(stat -c%s image.tar.gz 2>/dev/null || stat -f%z image.tar.gz)

curl -s -X POST \
  https://79x0as8g44.execute-api.us-east-1.amazonaws.com/prod/upload-url \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: qoXdZQNYbX1s4wnhmcBJG2APABlqjNVSao8CdM3j' \
  -d "{\"team_id\": \"AFM-xxxxxxxx\", \"file_size_bytes\": $FILE_SIZE}" \
  | tee /tmp/upload_resp.json | python3 -m json.tool

UPLOAD_URL=$(python3 -c "import json; print(json.load(open('/tmp/upload_resp.json'))['upload_url'])")

curl -X PUT "$UPLOAD_URL" \
  --upload-file image.tar.gz \
  --progress-bar \
  -w "\nHTTP Status: %{http_code}\n"

python3 - << 'EOF'
import json, os, subprocess, sys

resp    = json.load(open('/tmp/upload_resp.json'))
team_id = resp['s3_key'].split('/')[1]

etags_file = '/tmp/upload_etags.json'
if os.path.exists(etags_file):
    etags = json.load(open(etags_file))
    done  = {e['part_number'] for e in etags}
    print(f"Resuming — {len(done)} parts already done: {sorted(done)}")
else:
    etags = []
    done  = set()

print(f"Uploading {resp['num_parts']} parts ({resp['part_size']//1024//1024} MB each)...")

with open('image.tar.gz', 'rb') as f:
    for part in resp['part_urls']:
        n = part['part_number']
        if n in done:
            f.seek(resp['part_size'], 1)
            continue
        chunk = f.read(resp['part_size'])
        if not chunk:
            break
        tmp = f'/tmp/part_{n}.bin'
        with open(tmp, 'wb') as pf:
            pf.write(chunk)
        result = subprocess.run(
            ['curl', '-s', '-X', 'PUT', part['upload_url'],
             '--upload-file', tmp, '-D', '-', '-o', '/dev/null'],
            capture_output=True, text=True
        )
        os.remove(tmp)
        etag = ''
        for line in result.stdout.splitlines():
            if line.lower().startswith('etag:'):
                etag = line.split(':', 1)[1].strip().strip('\r').strip('"')
                break
        if not etag:
            print(f'  ERROR: No ETag for part {n}. Re-run to resume.')
            sys.exit(1)
        etags.append({'part_number': n, 'etag': etag})
        done.add(n)
        with open(etags_file, 'w') as ef:
            json.dump(etags, ef)
        print(f'  Part {n}/{resp["num_parts"]} done')

print(f'\nAll parts uploaded. Completing...')
body = json.dumps({
    'team_id':   team_id,
    's3_key':    resp['s3_key'],
    'upload_id': resp['upload_id'],
    'parts':     etags,
})
result = subprocess.run(
    ['curl', '-s', '-X', 'POST',
     'https://79x0as8g44.execute-api.us-east-1.amazonaws.com/prod/complete-upload',
     '-H', 'Content-Type: application/json',
     '-H', 'x-api-key: qoXdZQNYbX1s4wnhmcBJG2APABlqjNVSao8CdM3j',
     '-d', body],
    capture_output=True, text=True
)
print(json.dumps(json.loads(result.stdout), indent=2))
os.remove(etags_file)
EOF

S3_KEY=$(python3 -c "import json; print(json.load(open('/tmp/upload_resp.json'))['s3_key'])")

curl -s -X POST \
  https://79x0as8g44.execute-api.us-east-1.amazonaws.com/prod/submit \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: qoXdZQNYbX1s4wnhmcBJG2APABlqjNVSao8CdM3j' \
  -d "{\"team_id\": \"AFM-xxxxxxxx\", \"s3_key\": \"$S3_KEY\"}" \
  | python3 -m json.tool