[V1] V1 Enablement Oracle (#13726)

Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
Robert Shaw
2025-03-14 22:02:20 -07:00
committed by GitHub
co-authored by rshaw@neuralmagic.com <rshaw@neuralmagic.com> Nicolò Lucchesi Tyler Michael Smith Michael Goin
parent 8c0d15d5c5
commit d4d93db2c5
96 changed files with 1537 additions and 512 deletions
+8
View File
@@ -6,6 +6,13 @@ Run `pytest tests/samplers/test_beam_search.py`.
import pytest
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
"""We can run both engines for this test."""
pass
# FIXME(zhuohan): The test can not pass if we:
# 1. Increase max_tokens to 256.
# 2. Increase beam_width to 8.
@@ -15,6 +22,7 @@ BEAM_WIDTHS = [4]
MODELS = ["TinyLlama/TinyLlama-1.1B-Chat-v1.0"]
@pytest.mark.skip_v1 # FIXME: This fails on V1 right now.
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
+7
View File
@@ -8,6 +8,13 @@ import pytest
from vllm import SamplingParams
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
"""We can run both engines for this test."""
pass
# We also test with llama because it has generation_config to specify EOS
# (past regression).
MODELS = ["distilbert/distilgpt2", "meta-llama/Llama-3.2-1B"]
+8
View File
@@ -8,6 +8,14 @@ from vllm import SamplingParams
MODELS = ["distilbert/distilgpt2"]
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
This file tests V0 internals, so set VLLM_USE_V1=0.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_logits_processor_force_generate(
+9
View File
@@ -10,6 +10,15 @@ from ..conftest import VllmRunner
MODELS = ["distilbert/distilgpt2"]
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
This module is V0 only since it uses dtype=float, so
set VLLM_USE_V1=0 for all tests in the module.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype",
["float"]) # needed for comparing logprobs with HF
+7
View File
@@ -6,11 +6,18 @@ Run `pytest tests/samplers/test_no_bad_words.py`.
"""
from typing import Optional
import pytest
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
"""We can run both engines for this test."""
pass
def _generate(
model: LLM,
prompt: str,
+6
View File
@@ -7,6 +7,12 @@ from vllm import SamplingParams
MODELS = ["distilbert/distilgpt2"]
@pytest.fixture(autouse=True)
def v1(run_with_both_engines):
"""We can run both engines for this test."""
pass
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_ranks(
+9
View File
@@ -8,6 +8,15 @@ import torch.nn.functional as F
from vllm.model_executor.layers.rejection_sampler import RejectionSampler
from vllm.model_executor.utils import set_random_seed
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
This file tests V0 internals, so set VLLM_USE_V1=0.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
CUDA_DEVICES = [
f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
]
+8
View File
@@ -18,6 +18,14 @@ from vllm.sequence import SamplingParams, SequenceData, SequenceGroupMetadata
from vllm.utils import Counter, is_pin_memory_available
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
This file tests V0 internals, so set VLLM_USE_V1=0.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
class MockLogitsSampler(Sampler):
def __init__(self, fake_logits: torch.Tensor):
+3 -1
View File
@@ -17,7 +17,9 @@ RANDOM_SEEDS = list(range(5))
@pytest.fixture
def vllm_model(vllm_runner):
def vllm_model(vllm_runner, monkeypatch):
# This file relies on V0 internals.
monkeypatch.setenv("VLLM_USE_V1", "0")
with vllm_runner(MODEL, dtype="half") as vllm_model:
yield vllm_model
@@ -11,6 +11,14 @@ from vllm.model_executor.utils import set_random_seed
CUDA_DEVICES = [f"cuda:{i}" for i in range(1)]
@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
"""
This file tests V0 internals, so set VLLM_USE_V1=0.
"""
monkeypatch.setenv('VLLM_USE_V1', '0')
def get_zero_temperature_prob_dist(batch_size, k, vocab_size):
"""
Generates a fake temperature zero probability distribution.