@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 4caf7044e0
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 17, 04:15 UTC

Include tokens from prompt phase in `counter_generation_tokens` (#2802)

Ronen Schaffer committed Feb 23, 2024 at 00:00 UTC 4caf7044e052399f07089aa8f586d5bd641f7d53
3 files changed +39 -1
.buildkite/test-pipeline.yaml
+3
@@ -52,6 +52,9 @@ steps:
52 - label: LoRA Test
53 command: pytest -v -s lora
54
55 +- label: Metrics Test
56 + command: pytest -v -s metrics
57 +
58 - label: Benchmarks
59 working_dir: "/vllm-workspace/.buildkite"
60 commands:
tests/metrics/test_metrics.py
+33 -1
@@ -9,13 +9,16 @@ MODELS = [
9 @pytest.mark.parametrize("model", MODELS)
10 @pytest.mark.parametrize("dtype", ["float"])
11 @pytest.mark.parametrize("max_tokens", [128])
12 -def test_metrics(
12 +def test_metric_counter_prompt_tokens(
13 vllm_runner,
14 example_prompts,
15 model: str,
16 dtype: str,
17 max_tokens: int,
18 ) -> None:
19 + # Reset metric
20 + vllm.engine.metrics.counter_prompt_tokens.set_value({}, 0)
21 +
22 vllm_model = vllm_runner(model, dtype=dtype, disable_log_stats=False)
23 tokenizer = vllm_model.model.get_tokenizer()
24 prompt_token_counts = [len(tokenizer.encode(p)) for p in example_prompts]
@@ -31,3 +34,32 @@ def test_metrics(
34 assert vllm_prompt_token_count == metric_count, (
35 f"prompt token count: {vllm_prompt_token_count!r}\nmetric: {metric_count!r}"
36 )
37 +
38 +
39 +@pytest.mark.parametrize("model", MODELS)
40 +@pytest.mark.parametrize("dtype", ["float"])
41 +@pytest.mark.parametrize("max_tokens", [128])
42 +def test_metric_counter_generation_tokens(
43 + vllm_runner,
44 + example_prompts,
45 + model: str,
46 + dtype: str,
47 + max_tokens: int,
48 +) -> None:
49 + # Reset metric
50 + vllm.engine.metrics.counter_generation_tokens.set_value({}, 0)
51 +
52 + vllm_model = vllm_runner(model, dtype=dtype, disable_log_stats=False)
53 + vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
54 + tokenizer = vllm_model.model.get_tokenizer()
55 + metric_count = vllm.engine.metrics.counter_generation_tokens.get_value({})
56 + vllm_generation_count = 0
57 + for i in range(len(example_prompts)):
58 + vllm_output_ids, vllm_output_str = vllm_outputs[i]
59 + prompt_ids = tokenizer.encode(example_prompts[i])
60 + # vllm_output_ids contains both prompt tokens and generation tokens. We're interested only in the count of the generation tokens.
61 + vllm_generation_count += len(vllm_output_ids) - len(prompt_ids)
62 +
63 + assert vllm_generation_count == metric_count, (
64 + f"generation token count: {vllm_generation_count!r}\nmetric: {metric_count!r}"
65 + )
vllm/engine/llm_engine.py
+3
@@ -872,6 +872,9 @@ class LLMEngine:
872 num_prompt_tokens = sum(
873 len(seq_group.prompt_token_ids)
874 for seq_group in scheduler_outputs.scheduled_seq_groups)
875 + num_generation_tokens = sum(
876 + seq_group.num_seqs()
877 + for seq_group in scheduler_outputs.scheduled_seq_groups)
878 else:
879 num_generation_tokens = scheduler_outputs.num_batched_tokens
880