[CI] Shard (H200 MIG 35GB / MI355 DPX) Entrypoints Integration (Pooling) into named jobs (#58652)

Signed-off-by: Thang Nguyen <thangnguyenvn647@gmail.com>
Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: khluu <khluu000@gmail.com>
This commit is contained in:
Thang Nguyen
2026-09-28 00:21:08 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 khluu
parent 004e37ece2
commit 413ae4b167
22 changed files with 77 additions and 9 deletions
@@ -0,0 +1,111 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for ScoringIOProcessor post-tokenization helpers."""
from dataclasses import dataclass
from types import SimpleNamespace
import pytest
from tokenizers import Tokenizer, models, pre_tokenizers, processors
from transformers import TokenizersBackend
from vllm import TokensPrompt
from vllm.entrypoints.pooling.scoring.io_processor import (
CrossEncoderIOProcessor,
_apply_post_tokenization_to_token_type_ids,
)
from vllm.entrypoints.pooling.scoring.utils import compress_token_type_ids
from vllm.renderers import TokenizeParams
pytestmark = pytest.mark.skip_global_cleanup
@dataclass
class _DummyTokenizer:
truncation_side: str = "left"
# Outside the range of the prompt ids below, so a test can tell a pad
# token apart from a real one.
pad_token_id: int = 99999
@pytest.fixture
def llm_reranker_processor() -> CrossEncoderIOProcessor:
vocab = {
"[UNK]": 0,
"[CLS]": 1,
"[SEP]": 2,
"a": 3,
"bc": 4,
"##b": 5,
"##c": 6,
}
backend = Tokenizer(models.WordPiece(vocab, unk_token="[UNK]"))
backend.pre_tokenizer = pre_tokenizers.WhitespaceSplit()
backend.post_processor = processors.TemplateProcessing(
single="[CLS] $A [SEP]",
special_tokens=[("[CLS]", 1), ("[SEP]", 2)],
)
processor = CrossEncoderIOProcessor.__new__(CrossEncoderIOProcessor)
processor.model_config = SimpleNamespace(enable_prompt_embeds=False)
processor.tokenizer = TokenizersBackend(
tokenizer_object=backend,
unk_token="[UNK]",
cls_token="[CLS]",
sep_token="[SEP]",
)
processor.supports_score_template = False
processor.use_sep_token = False
processor.model = None
return processor
def test_llm_reranker_tokenization_is_independent_of_nonbinding_doc_limit(
llm_reranker_processor: CrossEncoderIOProcessor,
):
encode_kwargs = {"add_special_tokens": True}
_, uncapped = llm_reranker_processor.get_score_prompt("a", "bc", encode_kwargs)
_, nonbinding = llm_reranker_processor.get_score_prompt(
"a", "bc", encode_kwargs, max_tokens_per_doc=10
)
expected = [1, 3, 5, 6, 2]
assert uncapped["prompt_token_ids"] == expected
assert nonbinding["prompt_token_ids"] == expected
def test_token_type_ids_stay_aligned_with_a_truncated_padded_prompt():
"""The cross-encoder segment boundary must survive truncate + pad.
`token_type_ids` are parallel to `prompt_token_ids` and are reduced to a
single boundary index by `compress_token_type_ids`. If the two arrays are
truncated and padded in different orders they no longer describe the same
positions, and the model is told the query segment is empty.
"""
tokenizer = _DummyTokenizer()
num_query, num_doc = 20, 30
prompt = TokensPrompt(prompt_token_ids=list(range(num_query + num_doc)))
token_type_ids = [0] * num_query + [1] * num_doc
tok_params = TokenizeParams(
max_total_tokens=100,
pad_prompt_tokens=-1,
truncate_prompt_tokens=40,
truncation_side="left",
)
prompt = tok_params.apply_post_tokenization(tokenizer, prompt)
token_type_ids = _apply_post_tokenization_to_token_type_ids(
tokenizer, tok_params, token_type_ids
)
prompt_token_ids = prompt["prompt_token_ids"]
assert len(token_type_ids) == len(prompt_token_ids)
# Keeping the last 40 tokens drops the first 10 query tokens, so 10 query
# tokens survive and the document starts at index 10.
first_doc = compress_token_type_ids(token_type_ids)
assert first_doc == 10
assert prompt_token_ids[:first_doc] == list(range(10, num_query))
assert prompt_token_ids[first_doc:40] == list(range(num_query, 50))
@@ -0,0 +1,144 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for JinaRankingIOProcessor request building."""
from types import SimpleNamespace
from unittest.mock import MagicMock
import pytest
from tokenizers import Tokenizer, models, pre_tokenizers
from transformers import TokenizersBackend
from vllm import PoolingParams
from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor
from vllm.entrypoints.pooling.scoring.io_processor import JinaRankingIOProcessor
from vllm.entrypoints.pooling.scoring.protocol import RerankRequest
from vllm.entrypoints.pooling.scoring.typing import ScoringData
from vllm.entrypoints.pooling.typing import OfflineScoringInputsContext
from vllm.renderers import TokenizeParams
pytestmark = pytest.mark.skip_global_cleanup
def test_online_forwards_truncate_prompt_tokens_to_proxy(monkeypatch):
"""The proxy request handed to the base factory must carry
truncate_prompt_tokens/truncation_side from the real request.
JinaRankingIOProcessor swaps ctx.request for a proxy
PoolingCompletionRequest before delegating to the base factory, which
reads truncation off ctx.request. Dropping the fields on the proxy
silently disables truncate_prompt_tokens for Jina rerank/score.
"""
proc = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)
proc.valid_inputs_online = MagicMock(
return_value=ScoringData(data_1=["query"], data_2=["doc"])
)
proc._get_token_limits = MagicMock(return_value=(0, 0))
proc.ensure_str = MagicMock(side_effect=lambda data: list(data))
proc.format_docs_prompts_func = MagicMock(return_value="formatted prompt")
captured: dict[str, object] = {}
def _spy_base(self, ctx):
captured["truncate_prompt_tokens"] = ctx.request.truncate_prompt_tokens
captured["truncation_side"] = ctx.request.truncation_side
return []
monkeypatch.setattr(PoolingIOProcessor, "get_request_factory_online", _spy_base)
request = RerankRequest(
model="m",
query="query",
documents=["doc"],
truncate_prompt_tokens=512,
truncation_side="left",
)
ctx = MagicMock()
ctx.request = request
ctx.prompt_extras = None
proc.get_request_factory_online(ctx)
assert captured["truncate_prompt_tokens"] == 512
assert captured["truncation_side"] == "left"
# The real request is restored after delegating.
assert ctx.request is request
@pytest.fixture
def offline_processor_and_context():
backend = Tokenizer(models.WordLevel({"[UNK]": 0}, unk_token="[UNK]"))
backend.pre_tokenizer = pre_tokenizers.WhitespaceSplit()
proc = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)
proc.tokenizer = TokenizersBackend(tokenizer_object=backend)
proc.model_config = SimpleNamespace(max_model_len=1024, is_encoder_decoder=False)
proc.renderer = SimpleNamespace(
default_cmpl_tok_params=TokenizeParams(max_total_tokens=1024)
)
ctx = OfflineScoringInputsContext(
pooling_task="token_embed",
scoring_data=ScoringData(
data_1=["query alpha beta", "query gamma delta"],
data_2=["document one two", "document three four"],
),
pooling_params=PoolingParams(
extra_kwargs={
"chat_template_kwargs": {"instruction": "Keep this instruction"}
}
),
tokenization_kwargs=None,
chat_template=None,
lora_request=None,
priorities=None,
)
return proc, ctx
@pytest.mark.parametrize(
("query_limit", "doc_limit", "n_queries"),
[(None, None, 1), (0, 0, 2), (1, None, 1), (None, 2, 1), (1, 2, 1), (1, 2, 2)],
ids=["unset", "zero", "query", "doc", "both-1-to-n", "both-n-to-n"],
)
def test_offline_truncates_text_before_formatting(
offline_processor_and_context, query_limit, doc_limit, n_queries
):
proc, ctx = offline_processor_and_context
queries = ctx.scoring_data.data_1[:n_queries]
docs = ctx.scoring_data.data_2
ctx.scoring_data = ScoringData(data_1=queries, data_2=docs)
ctx.pooling_params.extra_kwargs.update(
{
name: limit
for name, limit in (
("max_tokens_per_query", query_limit),
("max_tokens_per_doc", doc_limit),
)
if limit is not None
}
)
factory, num_requests = proc.get_request_factory_offline(ctx)
requests = list(factory())
assert len(requests) == num_requests == n_queries
for i, request in enumerate(requests):
prompt = request["prompts"]["prompt"]
query = " ".join(queries[i].split()[: query_limit or None])
assert f"<query>\n{query}<|rerank_token|>\n</query>" in prompt
prompt_docs = docs if n_queries == 1 else [docs[i]]
for j, doc in enumerate(prompt_docs):
doc = " ".join(doc.split()[: doc_limit or None])
assert f'<passage id="{j}">\n{doc}<|embed_token|>\n</passage>' in prompt
assert prompt.count("<|embed_token|>") == len(prompt_docs)
assert prompt.count("<|rerank_token|>") == 1
assert "<instruct>\nKeep this instruction\n</instruct>" in prompt
@pytest.mark.parametrize("name", ["max_tokens_per_query", "max_tokens_per_doc"])
@pytest.mark.parametrize("limit", [-1, 1024])
def test_offline_rejects_invalid_token_limits(
offline_processor_and_context, name, limit
):
proc, ctx = offline_processor_and_context
ctx.pooling_params.extra_kwargs[name] = limit
with pytest.raises(ValueError, match=name):
proc.get_request_factory_offline(ctx)
@@ -0,0 +1,84 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from tokenizers import Tokenizer, models, pre_tokenizers
from transformers import TokenizersBackend
from vllm.entrypoints.pooling.scoring.typing import ScoreInput
from vllm.entrypoints.pooling.scoring.utils import (
truncate_text_to_tokens,
validate_score_input,
)
from vllm.exceptions import VLLMValidationError
@pytest.mark.parametrize(
("data_1", "data_2", "is_multimodal_model", "architecture", "message"),
[
pytest.param(
{"content": []},
"document",
False,
"TestModel",
"MultiModalParam is not supported for TestModel",
id="unsupported-multimodal-input",
),
pytest.param(
["query 1", "query 2"],
["document"],
False,
"TestModel",
"Input lengths must be either 1:1, 1:N or N:N",
id="incompatible-input-lengths",
),
pytest.param(
[],
["document"],
False,
"TestModel",
"At least one text element must be given",
id="empty-first-input",
),
pytest.param(
["query"],
[],
False,
"TestModel",
"At least one text_pair element must be given",
id="empty-second-input",
),
],
)
def test_validate_score_input_rejects_invalid_inputs(
data_1: ScoreInput | list[ScoreInput],
data_2: ScoreInput | list[ScoreInput],
is_multimodal_model: bool,
architecture: str,
message: str,
):
with pytest.raises(VLLMValidationError) as exc_info:
validate_score_input(
data_1,
data_2,
is_multimodal_model=is_multimodal_model,
architecture=architecture,
)
assert str(exc_info.value) == message
assert exc_info.value.parameter is None
assert exc_info.value.value is None
def test_truncate_text_to_tokens_handles_shared_character_offsets():
vocab = {
token: i for i, token in enumerate(sorted(pre_tokenizers.ByteLevel.alphabet()))
}
backend = Tokenizer(models.BPE(vocab=vocab, merges=[]))
backend.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
tokenizer = TokenizersBackend(tokenizer_object=backend)
truncated = truncate_text_to_tokens("aéx", tokenizer, max_tokens=2)
assert truncated == "a"
assert len(tokenizer(truncated, add_special_tokens=False)["input_ids"]) == 1