[Deprecation] Remove prompt_token_ids arg fallback in LLM.generate and LLM.embed (#18800)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-08-22 10:56:57 +08:00
committed by GitHub
parent 19fe1a0510
commit 8896eb72eb
24 changed files with 116 additions and 467 deletions
+6 -9
View File
@@ -18,10 +18,9 @@ def text_llm():
enforce_eager=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
@@ -88,10 +87,9 @@ def vision_llm():
seed=0,
)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
@@ -158,10 +156,9 @@ def thinking_llm():
seed=0,
)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+2 -3
View File
@@ -35,10 +35,9 @@ def llm():
enforce_eager=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+2 -3
View File
@@ -26,10 +26,9 @@ def llm():
enforce_eager=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+3 -49
View File
@@ -5,11 +5,9 @@ import weakref
import pytest
from vllm import LLM, PoolingParams, PoolingRequestOutput
from vllm import LLM, PoolingParams
from vllm.distributed import cleanup_dist_env_and_memory
from ...models.utils import check_embeddings_close
MODEL_NAME = "intfloat/multilingual-e5-small"
PROMPTS = [
@@ -48,57 +46,13 @@ def llm():
enforce_eager=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
def assert_outputs_match(o1: list[PoolingRequestOutput],
o2: list[PoolingRequestOutput]):
check_embeddings_close(
embeddings_0_lst=[o.outputs.data for o in o1],
embeddings_1_lst=[o.outputs.data for o in o2],
name_0="hf",
name_1="vllm",
tol=1e-2,
)
@pytest.mark.skip_global_cleanup
@pytest.mark.parametrize('prompt_token_ids', TOKEN_IDS)
def test_v1_v2_api_consistency_single_prompt_tokens(llm: LLM,
prompt_token_ids):
pooling_params = PoolingParams()
with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
v1_output = llm.encode(prompt_token_ids=prompt_token_ids,
pooling_params=pooling_params)
v2_output = llm.encode({"prompt_token_ids": prompt_token_ids},
pooling_params=pooling_params)
assert_outputs_match(v1_output, v2_output)
@pytest.mark.skip_global_cleanup
def test_v1_v2_api_consistency_multi_prompt_tokens(llm: LLM):
pooling_params = PoolingParams()
with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
v1_output = llm.encode(prompt_token_ids=TOKEN_IDS,
pooling_params=pooling_params)
v2_output = llm.encode(
[{
"prompt_token_ids": p
} for p in TOKEN_IDS],
pooling_params=pooling_params,
)
assert_outputs_match(v1_output, v2_output)
@pytest.mark.skip_global_cleanup
def test_multiple_pooling_params(llm: LLM):
pooling_params = [
+3 -40
View File
@@ -5,7 +5,7 @@ import weakref
import pytest
from vllm import LLM, RequestOutput, SamplingParams
from vllm import LLM, SamplingParams
from vllm.distributed import cleanup_dist_env_and_memory
MODEL_NAME = "distilbert/distilgpt2"
@@ -41,50 +41,13 @@ def llm():
gpu_memory_utilization=0.10,
enforce_eager=True)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
def assert_outputs_equal(o1: list[RequestOutput], o2: list[RequestOutput]):
assert [o.outputs for o in o1] == [o.outputs for o in o2]
@pytest.mark.skip_global_cleanup
@pytest.mark.parametrize('prompt_token_ids', TOKEN_IDS)
def test_v1_v2_api_consistency_single_prompt_tokens(llm: LLM,
prompt_token_ids):
sampling_params = SamplingParams(temperature=0.0, top_p=1.0)
with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
v1_output = llm.generate(prompt_token_ids=prompt_token_ids,
sampling_params=sampling_params)
v2_output = llm.generate({"prompt_token_ids": prompt_token_ids},
sampling_params=sampling_params)
assert_outputs_equal(v1_output, v2_output)
@pytest.mark.skip_global_cleanup
def test_v1_v2_api_consistency_multi_prompt_tokens(llm: LLM):
sampling_params = SamplingParams(temperature=0.0, top_p=1.0)
with pytest.warns(DeprecationWarning, match="'prompt_token_ids'"):
v1_output = llm.generate(prompt_token_ids=TOKEN_IDS,
sampling_params=sampling_params)
v2_output = llm.generate(
[{
"prompt_token_ids": p
} for p in TOKEN_IDS],
sampling_params=sampling_params,
)
assert_outputs_equal(v1_output, v2_output)
@pytest.mark.skip_global_cleanup
def test_multiple_sampling_params(llm: LLM):
sampling_params = [
@@ -48,10 +48,9 @@ def llm(request, monkeypatch_module):
max_num_seqs=128,
enforce_eager=True)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+2 -3
View File
@@ -36,10 +36,9 @@ def llm():
trust_remote_code=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+2 -3
View File
@@ -33,10 +33,9 @@ def llm():
enforce_eager=True,
seed=0)
with llm.deprecate_legacy_api():
yield weakref.proxy(llm)
yield weakref.proxy(llm)
del llm
del llm
cleanup_dist_env_and_memory()
+2 -4
View File
@@ -38,8 +38,7 @@ def test_model_load_and_run(vllm_runner, model_id: str, force_marlin: bool,
with vllm_runner(model_id) as llm:
# note: this does not test accuracy, just that we can run through
# see lm-eval tests for accuracy
outputs = llm.generate_greedy(prompts=["Hello my name is"],
max_tokens=10)
outputs = llm.generate_greedy(["Hello my name is"], max_tokens=10)
print(outputs[0][1])
@@ -90,8 +89,7 @@ def test_kv_cache_model_load_and_run(vllm_runner, model_id: str,
# note: this does not test accuracy, just that we can run through
# see lm-eval tests for accuracy
outputs = llm.generate_greedy(prompts=["Hello my name is"],
max_tokens=10)
outputs = llm.generate_greedy(["Hello my name is"], max_tokens=10)
print(outputs[0][1])
+1 -1
View File
@@ -46,5 +46,5 @@ def test_lm_head(
vllm_model.apply_model(check_model)
print(
vllm_model.generate_greedy(prompts=["Hello my name is"],
vllm_model.generate_greedy(["Hello my name is"],
max_tokens=10)[0][1])
@@ -127,13 +127,15 @@ def test_structured_output(
temperature=1.0,
max_tokens=4096,
guided_decoding=GuidedDecodingParams(json=sample_json_schema))
outputs = llm.generate(prompts=[
(f"Give an example JSON for an employee profile that fits this "
f"schema. Make the response as short as possible. Schema: "
f"{sample_json_schema}")
] * 2,
sampling_params=sampling_params,
use_tqdm=True)
prompt = ("Give an example JSON for an employee profile that fits this "
"schema. Make the response as short as possible. Schema: "
f"{sample_json_schema}")
outputs = llm.generate(
[prompt] * 2,
sampling_params=sampling_params,
use_tqdm=True,
)
assert outputs is not None
@@ -191,20 +193,24 @@ def test_structured_output(
with pytest.raises(ValueError,
match="The provided JSON schema contains features "
"not supported by xgrammar."):
prompt = (f"Give an example JSON for an employee profile that "
f"fits this schema: {unsupported_json_schema}. "
f"Make the response as short as possible.")
llm.generate(
prompts=[(f"Give an example JSON for an employee profile that "
f"fits this schema: {unsupported_json_schema}. "
f"Make the response as short as possible.")] * 2,
[prompt] * 2,
sampling_params=sampling_params,
use_tqdm=True)
use_tqdm=True,
)
else:
outputs = llm.generate(prompts=(
"Give an example JSON object for a grade "
"that fits this schema: "
f"{unsupported_json_schema}. Make the response as short as "
"possible."),
sampling_params=sampling_params,
use_tqdm=True)
prompt = (f"Give an example JSON object for a grade that "
f"fits this schema: {unsupported_json_schema}. "
f"Make the response as short as possible.")
outputs = llm.generate(
prompt,
sampling_params=sampling_params,
use_tqdm=True,
)
assert outputs is not None
for output in outputs:
assert output is not None
@@ -227,10 +233,9 @@ def test_structured_output(
max_tokens=1000,
guided_decoding=GuidedDecodingParams(grammar=sample_sql_ebnf))
outputs = llm.generate(
prompts=(
"Generate a sql statement that selects col_1 from "
"table_1 where it is equal to 1. Make the response as short as "
"possible."),
("Generate a sql statement that selects col_1 from "
"table_1 where it is equal to 1. Make the response as short as "
"possible."),
sampling_params=sampling_params,
use_tqdm=True,
)
@@ -261,10 +266,9 @@ def test_structured_output(
max_tokens=1000,
guided_decoding=GuidedDecodingParams(grammar=sample_sql_lark))
outputs = llm.generate(
prompts=(
"Generate a sql statement that selects col_1 from "
"table_1 where it is equal to 1. Make the response as short as "
"possible."),
("Generate a sql statement that selects col_1 from "
"table_1 where it is equal to 1. Make the response as short as "
"possible."),
sampling_params=sampling_params,
use_tqdm=True,
)
@@ -301,7 +305,6 @@ def test_structured_output(
guided_decoding=GuidedDecodingParams(grammar="not a grammar"))
with pytest.raises(ValueError, match="Failed to convert the grammar "):
llm.generate(
prompts=
("Generate a sql statement that selects col_1 from "
"table_1 where it is equal to 1. Make the response as short "
"as possible."),
@@ -316,11 +319,11 @@ def test_structured_output(
temperature=0.8,
top_p=0.95,
guided_decoding=GuidedDecodingParams(regex=sample_regex))
prompt = (f"Give an example IPv4 address with this regex: {sample_regex}. "
f"Make the response as short as possible.")
outputs = llm.generate(
prompts=[
(f"Give an example IPv4 address with this regex: {sample_regex}. "
f"Make the response as short as possible.")
] * 2,
[prompt] * 2,
sampling_params=sampling_params,
use_tqdm=True,
)
@@ -343,11 +346,13 @@ def test_structured_output(
temperature=0.8,
top_p=0.95,
guided_decoding=GuidedDecodingParams(choice=sample_guided_choice))
outputs = llm.generate(
prompts=("The best language for type-safe systems programming is "
"(Make the response as short as possible.) "),
("The best language for type-safe systems programming is "
"(Make the response as short as possible.) "),
sampling_params=sampling_params,
use_tqdm=True)
use_tqdm=True,
)
assert outputs is not None
for output in outputs:
assert output is not None
@@ -367,12 +372,14 @@ def test_structured_output(
temperature=1.0,
max_tokens=1000,
guided_decoding=GuidedDecodingParams(json=json_schema))
outputs = llm.generate(prompts=(
"Generate a JSON with the brand, model and car_type of the most "
"iconic car from the 90's. Make the response as short as "
"possible."),
sampling_params=sampling_params,
use_tqdm=True)
outputs = llm.generate(
("Generate a JSON with the brand, model and car_type of the most "
"iconic car from the 90's. Make the response as short as "
"possible."),
sampling_params=sampling_params,
use_tqdm=True,
)
assert outputs is not None
@@ -411,10 +418,11 @@ def test_structured_output(
guided_decoding=GuidedDecodingParams(json=json_schema))
outputs = llm.generate(
prompts=("Generate a description of a frog using 50 characters. "
"Make the response as short as possible."),
("Generate a description of a frog using 50 characters. "
"Make the response as short as possible."),
sampling_params=sampling_params,
use_tqdm=True)
use_tqdm=True,
)
assert outputs is not None
@@ -498,7 +506,7 @@ Make the response as short as possible.
"""
# Change this once other backends support structural_tag
outputs = llm.generate(prompts=prompt,
outputs = llm.generate(prompt,
sampling_params=sampling_params,
use_tqdm=True)
assert outputs is not None
@@ -639,15 +647,13 @@ def test_structured_output_auto_mode(
f"{unsupported_json_schema}. Make the response as short as possible.")
# This would fail with the default of "xgrammar", but in "auto"
# we will handle fallback automatically.
outputs = llm.generate(prompts=prompts,
outputs = llm.generate(prompts,
sampling_params=sampling_params,
use_tqdm=True)
# Make sure `auto` backend handling doesn't mess up sampling_params
# and that we can reuse it without error.
outputs.extend(
llm.generate(prompts=prompts,
sampling_params=sampling_params,
use_tqdm=True))
llm.generate(prompts, sampling_params=sampling_params, use_tqdm=True))
assert outputs is not None
for output in outputs:
@@ -705,7 +711,7 @@ def test_guidance_no_additional_properties(monkeypatch: pytest.MonkeyPatch):
max_tokens=256,
guided_decoding=guided_params)
outputs = llm.generate(prompts=prompt, sampling_params=sampling_params)
outputs = llm.generate(prompt, sampling_params=sampling_params)
assert outputs is not None
generated_text = outputs[0].outputs[0].text
assert generated_text is not None