Include tokens from prompt phase in counter_generation_tokens

tests/conftest.py

@@ -165,13 +165,15 @@ def __init__(
         model_name: str,
         tokenizer_name: Optional[str] = None,
         dtype: str = "half",
+        disable_log_stats: bool = True,
     ) -> None:
         self.model = LLM(
             model=model_name,
             tokenizer=tokenizer_name,
             trust_remote_code=True,
             dtype=dtype,
             swap_space=0,
+            disable_log_stats=disable_log_stats,
         )
 
     def generate(

tests/metrics/test_metrics.py

@@ -0,0 +1,32 @@
+import pytest
+import vllm.engine.metrics
+
+MODELS = [
+    "facebook/opt-125m",
+]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["float"])
+@pytest.mark.parametrize("max_tokens", [128])
+def test_metrics(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+    max_tokens: int,
+) -> None:
+    vllm_model = vllm_runner(model, dtype=dtype, disable_log_stats=False)
+    vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+    tokenizer = vllm_model.model.get_tokenizer()
+    metric_count = vllm.engine.metrics.counter_generation_tokens.get_value({})
+    vllm_generation_count = 0
+    for i in range(len(example_prompts)):
+        vllm_output_ids, vllm_output_str = vllm_outputs[i]
+        prompt_ids = tokenizer.encode(example_prompts[i])
+        # vllm_output_ids contains both prompt tokens and generation tokens. We're interested only in the count of the generation tokens.
+        vllm_generation_count += len(vllm_output_ids) - len(prompt_ids)
+
+    assert vllm_generation_count == metric_count, (
+        f"generation token count: {vllm_generation_count!r}\nmetric: {metric_count!r}"
+    )

vllm/engine/llm_engine.py

@@ -854,6 +854,10 @@ def _get_stats(self,
             # Number of Tokens.
             if prompt_run:
                 num_prompt_tokens = scheduler_outputs.num_batched_tokens
+                num_generation_tokens = sum([
+                    seq_group.num_seqs()
+                    for seq_group in scheduler_outputs.scheduled_seq_groups
+                ])
             else:
                 num_generation_tokens = scheduler_outputs.num_batched_tokens