biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V0 Deprecation] Remove LLMEngine (#25033)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-09-20 17:56:30 -07:00
committed by GitHub
parent 367a480bd3
commit 52c2a8d4ad
29 changed files with 65 additions and 2763 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
tests/tokenization/test_detokenize.py
View File

@@ -352,58 +352,3 @@ def test_decode_prompt_logprobs(complete_sequence: str,
logprobs[token_id + 1].decoded_token
for token_id, logprobs in zip(token_ids[1:], decoded_prompt_logprobs)
])
@pytest.mark.parametrize("model", ["facebook/opt-125m"])
@pytest.mark.parametrize("chunked_prefill_token_size", [1, 4, 7, 16, -1])
def test_decode_prompt_logprobs_chunked_prefill(
vllm_runner,
model,
chunked_prefill_token_size: int,
example_prompts,
monkeypatch,
):
# VLLM V1 does not use incremental detokenization for
# prompt logprobs, so this test strategy is irrelevant.
monkeypatch.setenv("VLLM_USE_V1", "0")
max_num_seqs = 256
enable_chunked_prefill = False
max_num_batched_tokens = None
if chunked_prefill_token_size != -1:
enable_chunked_prefill = True
max_num_seqs = min(chunked_prefill_token_size, max_num_seqs)
max_num_batched_tokens = chunked_prefill_token_size
with vllm_runner(model,
dtype="half",
max_logprobs=5,
gpu_memory_utilization=0.5,
enable_chunked_prefill=enable_chunked_prefill,
max_num_batched_tokens=max_num_batched_tokens,
max_num_seqs=max_num_seqs) as vllm_model:
vllm_sampling_params = SamplingParams(max_tokens=10,
logprobs=5,
prompt_logprobs=5,
temperature=0.0)
vllm_results = vllm_model.llm.generate(
example_prompts, sampling_params=vllm_sampling_params)
for idx, result in enumerate(vllm_results):
assert result.prompt_logprobs is not None
assert result.prompt_logprobs[0] is None
# Compared detokenized prompts ids to original prompt.
generated_string = ""
for (prompt_token,
prompt_logprobs) in zip(result.prompt_token_ids[1:],
result.prompt_logprobs[1:]):
# prompt_logprobs is a dict of the token_id: logprob
# We select the token_id corresponding to the actual prompt
# Decoded token in the detokenized string corresponding to this
# prompt token.
generated_string += prompt_logprobs[prompt_token].decoded_token
assert generated_string == example_prompts[idx], (
"Detokenized prompt logprobs do not match original prompt")