[Core][Metrics][BugFix] Replace num_cached_tokens/num_external_computed_tokens with PrefillStats (#37460)

Related to `Counters can only be incremented by non-negative amounts`
error with the `vllm:prompt_tokens_by_source_total` metric.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Co-authored-by: Or Ozeri <or@ozery.com>
This commit is contained in:
Mark McLoughlin
2026-04-14 09:00:45 +01:00
committed by GitHub
co-authored by Or Ozeri
parent 25b3242d8b
commit d3af8c1831
9 changed files with 187 additions and 102 deletions
+10 -4
View File
@@ -84,6 +84,7 @@ def test_incremental_detokenization(
engine_core = MockEngineCore(
tokens_list=dummy_test_vectors.generation_tokens,
prompts_list=dummy_test_vectors.prompt_tokens,
request_ids=[req.request_id for req in requests],
)
@@ -506,6 +507,7 @@ def test_logprobs_processor(
engine_core = MockEngineCore(
tokens_list=dummy_test_vectors.generation_tokens,
prompts_list=dummy_test_vectors.prompt_tokens,
generated_logprobs_raw=None
if num_sample_logprobs is None
else dummy_test_vectors.generation_logprobs,
@@ -691,6 +693,7 @@ def test_stop_token(
engine_core = MockEngineCore(
tokens_list=[generation_tokens],
prompts_list=dummy_test_vectors.prompt_tokens,
generated_logprobs_raw=[generation_logprobs] if do_logprobs else None,
prompt_logprobs_raw=None,
eos_token_id=sampling_params.eos_token_id,
@@ -794,6 +797,7 @@ def test_stop_string(
engine_core = MockEngineCore(
tokens_list=dummy_test_vectors.generation_tokens,
prompts_list=dummy_test_vectors.prompt_tokens,
generated_logprobs_raw=dummy_test_vectors.generation_logprobs
if num_sample_logprobs
else None,
@@ -917,6 +921,7 @@ def test_iteration_stats(dummy_test_vectors):
engine_core = MockEngineCore(
dummy_test_vectors.generation_tokens,
dummy_test_vectors.prompt_tokens,
request_ids=[req.request_id for req in requests],
)
@@ -927,7 +932,7 @@ def test_iteration_stats(dummy_test_vectors):
inactive_request = requests[num_active]
# First iteration has 2 prefills.
outputs = engine_core.get_outputs()[:num_active]
outputs = engine_core.get_outputs(num_active)
iteration_stats = IterationStats()
output_processor.process_outputs(outputs, engine_core_timestamp, iteration_stats)
total_prompt_tokens = sum(
@@ -941,7 +946,7 @@ def test_iteration_stats(dummy_test_vectors):
assert iteration_stats.num_generation_tokens == num_active
# Just decodes in this step.
outputs = engine_core.get_outputs()[:num_active]
outputs = engine_core.get_outputs(num_active)
iteration_stats = IterationStats()
output_processor.process_outputs(outputs, engine_core_timestamp, iteration_stats)
@@ -951,7 +956,7 @@ def test_iteration_stats(dummy_test_vectors):
# Add a new request - prefill and 2 decodes in this step.
output_processor.add_request(inactive_request, None)
num_active += 1
outputs = engine_core.get_outputs()[:num_active]
outputs = engine_core.get_outputs(num_active)
iteration_stats = IterationStats()
output_processor.process_outputs(outputs, engine_core_timestamp, iteration_stats)
total_prompt_tokens = len(dummy_test_vectors.prompt_tokens[num_active - 1])
@@ -960,7 +965,7 @@ def test_iteration_stats(dummy_test_vectors):
assert iteration_stats.num_generation_tokens == num_active
# Just decodes in this step.
outputs = engine_core.get_outputs()[:num_active]
outputs = engine_core.get_outputs(num_active)
iteration_stats = IterationStats()
output_processor.process_outputs(outputs, engine_core_timestamp, iteration_stats)
@@ -1003,6 +1008,7 @@ def test_lora_request_tracking(log_stats: bool, dummy_test_vectors):
engine_core = MockEngineCore(
dummy_test_vectors.generation_tokens,
dummy_test_vectors.prompt_tokens,
request_ids=[req.request_id for req in requests],
)