| 1 | #!/usr/bin/env bash |
| 2 | set -euo pipefail |
| 3 | |
| 4 | repo_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../../.." && pwd)" |
| 5 | |
| 6 | model="ibm-research/PowerMoE-3b" |
| 7 | port=8000 |
| 8 | num_decode_tokens=128 |
| 9 | server_pid= |
| 10 | server_log= |
| 11 | |
| 12 | # XPU MoE kernels may not safely handle the -1 routes emitted for padding. |
| 13 | export VLLM_MOE_SKIP_PADDING=0 |
| 14 | |
| 15 | cd "$repo_root" |
| 16 | |
| 17 | cleanup_server() { |
| 18 | local attempt |
| 19 | |
| 20 | if [[ -n "$server_pid" ]] && kill -0 "$server_pid" 2>/dev/null; then |
| 21 | kill "$server_pid" 2>/dev/null || true |
| 22 | |
| 23 | for ((attempt = 0; attempt < 30; attempt++)); do |
| 24 | if ! kill -0 "$server_pid" 2>/dev/null; then |
| 25 | break |
| 26 | fi |
| 27 | sleep 1 |
| 28 | done |
| 29 | |
| 30 | if kill -0 "$server_pid" 2>/dev/null; then |
| 31 | echo "vLLM did not stop after 30 seconds; sending SIGKILL" >&2 |
| 32 | kill -KILL "$server_pid" 2>/dev/null || true |
| 33 | fi |
| 34 | |
| 35 | wait "$server_pid" 2>/dev/null || true |
| 36 | fi |
| 37 | |
| 38 | server_pid= |
| 39 | } |
| 40 | |
| 41 | on_exit() { |
| 42 | local status=$? |
| 43 | |
| 44 | cleanup_server |
| 45 | if [[ "$status" -ne 0 && -n "$server_log" && -f "$server_log" ]]; then |
| 46 | if command -v buildkite-agent >/dev/null 2>&1; then |
| 47 | buildkite-agent artifact upload "$server_log" || |
| 48 | echo "Failed to upload $server_log as a Buildkite artifact" >&2 |
| 49 | fi |
| 50 | fi |
| 51 | [[ -z "$server_log" ]] || rm -f "$server_log" |
| 52 | exit "$status" |
| 53 | } |
| 54 | trap on_exit EXIT |
| 55 | |
| 56 | wait_for_server() { |
| 57 | for _ in $(seq 1 600); do |
| 58 | if curl -sf "http://127.0.0.1:$port/health" >/dev/null; then |
| 59 | return |
| 60 | fi |
| 61 | if ! kill -0 "$server_pid" 2>/dev/null; then |
| 62 | echo "vLLM exited before becoming healthy; see the server log above." >&2 |
| 63 | return 1 |
| 64 | fi |
| 65 | sleep 1 |
| 66 | done |
| 67 | echo "Timed out waiting for vLLM server health check" >&2 |
| 68 | return 1 |
| 69 | } |
| 70 | |
| 71 | request_outputs() { |
| 72 | local prompts_json=$1 |
| 73 | local expected_choices=$2 |
| 74 | local response |
| 75 | |
| 76 | response=$(printf \ |
| 77 | '{"model":"%s","prompt":%s,"temperature":0.7,"top_p":0.95,"seed":0,"max_tokens":%s,"logprobs":64,"ignore_eos":true,"return_token_ids":true}' \ |
| 78 | "$model" "$prompts_json" "$num_decode_tokens" | |
| 79 | curl --fail --silent --show-error \ |
| 80 | --max-time 1200 \ |
| 81 | "http://127.0.0.1:$port/v1/completions" \ |
| 82 | -H "Content-Type: application/json" \ |
| 83 | --data-binary @-) |
| 84 | |
| 85 | local -a token_id_arrays logprob_arrays |
| 86 | mapfile -t token_id_arrays < <( |
| 87 | grep -o '"token_ids":\[[^]]*\]' <<<"$response" | |
| 88 | sed 's/^"token_ids":\[//; s/\]$//' |
| 89 | ) |
| 90 | mapfile -t logprob_arrays < <( |
| 91 | grep -o '"token_logprobs":\[[^]]*\]' <<<"$response" | |
| 92 | sed 's/^"token_logprobs":\[//; s/\]$//' |
| 93 | ) |
| 94 | if [[ "${#token_id_arrays[@]}" -ne "$expected_choices" ]] || |
| 95 | [[ "${#logprob_arrays[@]}" -ne "$expected_choices" ]]; then |
| 96 | echo "Expected $expected_choices choices with token IDs and logits" >&2 |
| 97 | return 1 |
| 98 | fi |
| 99 | local index |
| 100 | local array |
| 101 | local value |
| 102 | local -a values |
| 103 | for index in "${!logprob_arrays[@]}"; do |
| 104 | array="${logprob_arrays[index]}" |
| 105 | IFS=, read -r -a values <<<"$array" |
| 106 | if [[ "${#values[@]}" -ne "$num_decode_tokens" ]]; then |
| 107 | echo "Completion returned an unexpected number of decode steps" >&2 |
| 108 | return 1 |
| 109 | fi |
| 110 | IFS=, read -r -a values <<<"${token_id_arrays[index]}" |
| 111 | if [[ "${#values[@]}" -ne "$num_decode_tokens" ]]; then |
| 112 | echo "Completion returned an unexpected number of generated tokens" >&2 |
| 113 | return 1 |
| 114 | fi |
| 115 | for value in "${values[@]}"; do |
| 116 | if [[ ! "$value" =~ ^[0-9]+$ ]]; then |
| 117 | echo "Completion returned an invalid token ID" >&2 |
| 118 | return 1 |
| 119 | fi |
| 120 | done |
| 121 | IFS=, read -r -a values <<<"$array" |
| 122 | for value in "${values[@]}"; do |
| 123 | if [[ ! "$value" =~ ^-?([0-9]+([.][0-9]*)?|[.][0-9]+)([eE][+-]?[0-9]+)?$ ]]; then |
| 124 | echo "Completion returned non-finite raw logits" >&2 |
| 125 | return 1 |
| 126 | fi |
| 127 | done |
| 128 | printf '%s|%s\n' "${token_id_arrays[index]}" "${logprob_arrays[index]}" |
| 129 | done |
| 130 | } |
| 131 | |
| 132 | make_prompt() { |
| 133 | local length=$1 |
| 134 | local offset=${2:-0} |
| 135 | local i |
| 136 | local prompt="Use the following system-design context to answer the question. Scenario $offset:" |
| 137 | for ((i = 0; i < length; i++)); do |
| 138 | prompt+=" context" |
| 139 | done |
| 140 | prompt+="\\n\\nQuestion: How should batch-invariant inference behave when this same request is decoded alongside unrelated requests?\\nAnswer:" |
| 141 | printf '"%s"' "$prompt" |
| 142 | } |
| 143 | |
| 144 | first_difference() { |
| 145 | local actual=$1 |
| 146 | local expected=$2 |
| 147 | local position=$3 |
| 148 | local actual_tokens=${actual%%|*} |
| 149 | local expected_tokens=${expected%%|*} |
| 150 | local actual_logits=${actual#*|} |
| 151 | local expected_logits=${expected#*|} |
| 152 | local -a actual_token_values |
| 153 | local -a expected_token_values |
| 154 | local -a actual_logit_values |
| 155 | local -a expected_logit_values |
| 156 | local step |
| 157 | |
| 158 | IFS=, read -r -a actual_token_values <<<"$actual_tokens" |
| 159 | IFS=, read -r -a expected_token_values <<<"$expected_tokens" |
| 160 | IFS=, read -r -a actual_logit_values <<<"$actual_logits" |
| 161 | IFS=, read -r -a expected_logit_values <<<"$expected_logits" |
| 162 | for ((step = 0; step < num_decode_tokens; step++)); do |
| 163 | if [[ "${actual_token_values[step]}" != "${expected_token_values[step]}" ]]; then |
| 164 | printf 'batch position %s: step=%s, baseline token=%s, actual token=%s' \ |
| 165 | "$position" \ |
| 166 | "$step" \ |
| 167 | "${expected_token_values[step]}" \ |
| 168 | "${actual_token_values[step]}" |
| 169 | return |
| 170 | fi |
| 171 | if [[ "${actual_logit_values[step]}" != "${expected_logit_values[step]}" ]]; then |
| 172 | printf 'batch position %s: step=%s, token=%s, baseline logit=%s, actual logit=%s' \ |
| 173 | "$position" \ |
| 174 | "$step" \ |
| 175 | "${expected_token_values[step]}" \ |
| 176 | "${expected_logit_values[step]}" \ |
| 177 | "${actual_logit_values[step]}" |
| 178 | return |
| 179 | fi |
| 180 | done |
| 181 | } |
| 182 | |
| 183 | run_e2e() { |
| 184 | local needle |
| 185 | local batch |
| 186 | local baseline_output |
| 187 | local batch_response |
| 188 | local difference |
| 189 | local position |
| 190 | local -a fillers |
| 191 | local -a batch_outputs |
| 192 | local -a differences=() |
| 193 | |
| 194 | echo "Testing PowerMoE-3b: TP=2, VLLM_BATCH_INVARIANT=1, VLLM_MOE_SKIP_PADDING=0" |
| 195 | server_log=$(mktemp "${TMPDIR:-/tmp}/xpu-batch-invariance.XXXXXX.log") |
| 196 | VLLM_BATCH_INVARIANT=1 vllm serve "$model" \ |
| 197 | --tensor-parallel-size 2 \ |
| 198 | --max-model-len 1024 \ |
| 199 | --max-num-seqs 16 \ |
| 200 | --max-num-batched-tokens 512 \ |
| 201 | --enable-chunked-prefill \ |
| 202 | --no-enable-prefix-caching \ |
| 203 | --enforce-eager \ |
| 204 | --seed 0 \ |
| 205 | --max-logprobs 64 \ |
| 206 | --logprobs-mode raw_logits \ |
| 207 | --port "$port" > >(tee "$server_log") 2>&1 & |
| 208 | server_pid=$! |
| 209 | wait_for_server |
| 210 | |
| 211 | needle=$(make_prompt 257) |
| 212 | local filler_lengths=(7 31 63 95 127 159 191 223 255 287 319 383 447) |
| 213 | for position in "${!filler_lengths[@]}"; do |
| 214 | fillers+=( "$(make_prompt "${filler_lengths[position]}" "$((position + 1))")" ) |
| 215 | done |
| 216 | batch="[$needle,${fillers[0]},${fillers[1]},${fillers[2]},${fillers[3]},${fillers[4]},${fillers[5]},$needle,${fillers[6]},${fillers[7]},${fillers[8]},${fillers[9]},${fillers[10]},${fillers[11]},${fillers[12]},$needle]" |
| 217 | |
| 218 | echo "Requesting BS=1 baseline ($num_decode_tokens decode tokens)..." |
| 219 | baseline_output=$(request_outputs "$needle" 1) |
| 220 | echo "Received BS=1 baseline." |
| 221 | echo "Requesting BS=16 batch ($((16 * num_decode_tokens)) decode tokens)..." |
| 222 | batch_response=$(request_outputs "$batch" 16) |
| 223 | mapfile -t batch_outputs <<<"$batch_response" |
| 224 | echo "Received BS=16 batch; comparing generated tokens and logits..." |
| 225 | |
| 226 | for position in 0 7 15; do |
| 227 | difference=$(first_difference "${batch_outputs[position]}" "$baseline_output" "$position") |
| 228 | if [[ -n "$difference" ]]; then |
| 229 | differences+=("$difference") |
| 230 | fi |
| 231 | done |
| 232 | |
| 233 | if [[ "${#differences[@]}" -ne 0 ]]; then |
| 234 | echo "Batch-invariant PowerMoE-3b output changed: ${differences[*]}" >&2 |
| 235 | return 1 |
| 236 | fi |
| 237 | echo "PowerMoE-3b batch invariance check passed" |
| 238 | cleanup_server |
| 239 | rm -f "$server_log" |
| 240 | server_log= |
| 241 | } |
| 242 | |
| 243 | run_e2e |