[Renderer] Separate out RendererConfig from ModelConfig (#30145)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
@@ -16,6 +16,7 @@ from vllm.config import (
|
||||
LoadConfig,
|
||||
ModelConfig,
|
||||
ParallelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
@@ -216,6 +217,7 @@ def create_vllm_config(
|
||||
|
||||
return VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
parallel_config=parallel_config,
|
||||
scheduler_config=scheduler_config,
|
||||
|
||||
@@ -8,7 +8,7 @@ import pytest
|
||||
import torch
|
||||
|
||||
import vllm.v1.core.kv_cache_utils as kv_cache_utils
|
||||
from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.config import ModelConfig, RendererConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.lora.request import LoRARequest
|
||||
from vllm.multimodal.inputs import (
|
||||
MultiModalFeatureSpec,
|
||||
@@ -667,7 +667,10 @@ def test_metrics_empty_stats():
|
||||
|
||||
def test_get_kv_cache_configs_multiple_workers():
|
||||
model_config = ModelConfig(max_model_len=16)
|
||||
vllm_config = VllmConfig(model_config=model_config)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
)
|
||||
|
||||
ref_kv_cache_spec = new_kv_cache_spec()
|
||||
same_kv_cache_specs = [
|
||||
@@ -1136,6 +1139,7 @@ def test_estimate_max_model_len(model_id, max_model_len, want_estimated_max_len)
|
||||
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
scheduler_config=scheduler_config,
|
||||
)
|
||||
|
||||
@@ -1175,6 +1179,7 @@ def test_get_max_concurrency_for_kv_cache_config():
|
||||
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
scheduler_config=scheduler_config,
|
||||
)
|
||||
|
||||
@@ -1293,7 +1298,10 @@ def test_allocate_with_lookahead():
|
||||
def test_get_kv_cache_config_one_worker():
|
||||
# pass max_model_len to pass check_enough_kv_cache_memory
|
||||
model_config = ModelConfig(max_model_len=16)
|
||||
vllm_config = VllmConfig(model_config=model_config)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
)
|
||||
|
||||
mem_per_block_per_layer = 16 * 2 * 64 * 4 * 2
|
||||
# all layers are full attention -> single group
|
||||
@@ -1584,7 +1592,11 @@ def test_get_kv_cache_config_one_worker():
|
||||
|
||||
def test_get_kv_cache_configs_attention_free():
|
||||
kv_cache_specs: dict[str, KVCacheSpec] = {}
|
||||
vllm_config = VllmConfig(model_config=ModelConfig(max_model_len=16))
|
||||
model_config = ModelConfig(max_model_len=16)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
)
|
||||
kv_cache_configs = get_kv_cache_configs(vllm_config, [kv_cache_specs], [0])
|
||||
assert kv_cache_configs == [
|
||||
KVCacheConfig(
|
||||
|
||||
@@ -11,6 +11,7 @@ from vllm.config import (
|
||||
ECTransferConfig,
|
||||
KVTransferConfig,
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
@@ -1563,6 +1564,7 @@ def create_scheduler_with_priority(
|
||||
vllm_config = VllmConfig(
|
||||
scheduler_config=scheduler_config,
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
kv_transfer_config=kv_transfer_config,
|
||||
speculative_config=speculative_config,
|
||||
|
||||
@@ -9,6 +9,7 @@ from vllm.config import (
|
||||
ECTransferConfig,
|
||||
KVTransferConfig,
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
@@ -132,6 +133,7 @@ def create_scheduler(
|
||||
vllm_config = VllmConfig(
|
||||
scheduler_config=scheduler_config,
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
kv_transfer_config=kv_transfer_config,
|
||||
speculative_config=speculative_config,
|
||||
|
||||
@@ -15,6 +15,7 @@ from vllm.config import (
|
||||
ECTransferConfig,
|
||||
KVTransferConfig,
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
@@ -522,6 +523,7 @@ def test_encoder_instance_zero_kv_cache(
|
||||
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
scheduler_config=scheduler_config,
|
||||
kv_transfer_config=kv_transfer_config,
|
||||
|
||||
@@ -5,7 +5,14 @@ import pytest
|
||||
|
||||
from vllm.assets.image import ImageAsset
|
||||
from vllm.assets.video import VideoAsset
|
||||
from vllm.config import CacheConfig, DeviceConfig, ModelConfig, VllmConfig
|
||||
from vllm.config import (
|
||||
CacheConfig,
|
||||
DeviceConfig,
|
||||
ModelConfig,
|
||||
MultiModalConfig,
|
||||
RendererConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
from vllm.sampling_params import SamplingParams
|
||||
from vllm.v1.engine import input_processor as input_processor_mod
|
||||
from vllm.v1.engine.input_processor import InputProcessor
|
||||
@@ -44,22 +51,21 @@ def _mock_input_processor(
|
||||
monkeypatch.setattr(VllmConfig, "__post_init__", lambda self: None, raising=True)
|
||||
|
||||
model_config = ModelConfig(
|
||||
skip_tokenizer_init=True,
|
||||
max_model_len=128,
|
||||
mm_processor_cache_gb=mm_cache_gb,
|
||||
generation_config="vllm",
|
||||
)
|
||||
model_config.multimodal_config = MultiModalConfig(mm_processor_cache_gb=mm_cache_gb)
|
||||
|
||||
renderer_config = RendererConfig(
|
||||
model_config=model_config,
|
||||
tokenizer="dummy",
|
||||
skip_tokenizer_init=True,
|
||||
)
|
||||
|
||||
# Minimal multimodal_config to satisfy references in
|
||||
# Processor.process_inputs.
|
||||
class _MockMMConfig:
|
||||
def __init__(self, gb: float):
|
||||
self.mm_processor_cache_gb = gb
|
||||
|
||||
model_config.multimodal_config = _MockMMConfig(mm_cache_gb) # type: ignore[attr-defined]
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=renderer_config,
|
||||
cache_config=CacheConfig(enable_prefix_caching=enable_prefix_caching),
|
||||
device_config=DeviceConfig(device="cpu"),
|
||||
)
|
||||
|
||||
@@ -15,6 +15,7 @@ from vllm.config import (
|
||||
DeviceConfig,
|
||||
KVTransferConfig,
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
@@ -127,6 +128,7 @@ def create_vllm_config(
|
||||
return VllmConfig(
|
||||
scheduler_config=scheduler_config,
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
kv_transfer_config=kv_transfer_config,
|
||||
device_config=DeviceConfig("cpu"),
|
||||
|
||||
@@ -19,6 +19,7 @@ from vllm.config import (
|
||||
DeviceConfig,
|
||||
ModelConfig,
|
||||
ParallelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
@@ -61,6 +62,7 @@ def _create_proposer(
|
||||
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=CacheConfig(),
|
||||
speculative_config=speculative_config,
|
||||
device_config=DeviceConfig(device=current_platform.device_type),
|
||||
|
||||
@@ -18,6 +18,7 @@ from vllm.config import (
|
||||
DeviceConfig,
|
||||
ModelConfig,
|
||||
ParallelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
@@ -46,6 +47,7 @@ def _create_mtp_proposer(num_speculative_tokens: int) -> EagleProposer:
|
||||
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=CacheConfig(),
|
||||
speculative_config=speculative_config,
|
||||
device_config=DeviceConfig(device=current_platform.device_type),
|
||||
|
||||
@@ -4,6 +4,7 @@ import numpy as np
|
||||
|
||||
from vllm.config import (
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SpeculativeConfig,
|
||||
VllmConfig,
|
||||
)
|
||||
@@ -69,6 +70,7 @@ def test_ngram_proposer():
|
||||
return NgramProposer(
|
||||
vllm_config=VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
speculative_config=SpeculativeConfig(
|
||||
prompt_lookup_min=min_n,
|
||||
prompt_lookup_max=max_n,
|
||||
|
||||
@@ -6,7 +6,7 @@ from concurrent.futures import Future
|
||||
import pytest
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
from vllm.config import StructuredOutputsConfig, VllmConfig
|
||||
from vllm.config import RendererConfig, StructuredOutputsConfig, VllmConfig
|
||||
from vllm.config.model import ModelConfig
|
||||
from vllm.config.parallel import ParallelConfig
|
||||
from vllm.config.speculative import SpeculativeConfig
|
||||
@@ -72,8 +72,11 @@ def test_backend_guidance_rollback_terminated():
|
||||
def test_grammar_bitmask_with_specdec():
|
||||
tokenizer = AutoTokenizer.from_pretrained(TOKENIZER)
|
||||
prompt = tokenizer.encode('{"a": "b"}')
|
||||
|
||||
model_config = ModelConfig(tokenizer=TOKENIZER)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=ModelConfig(tokenizer=TOKENIZER),
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config, tokenizer=TOKENIZER),
|
||||
structured_outputs_config=StructuredOutputsConfig(backend="guidance"),
|
||||
speculative_config=SpeculativeConfig(model="[ngram]", num_speculative_tokens=3),
|
||||
)
|
||||
@@ -137,8 +140,11 @@ def test_grammar_init_async_and_sync(async_grammar):
|
||||
|
||||
# Use "external_launcher" for sync mode, None for async mode
|
||||
executor_backend = None if async_grammar else "external_launcher"
|
||||
|
||||
model_config = ModelConfig(tokenizer=TOKENIZER)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=ModelConfig(tokenizer=TOKENIZER),
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config, tokenizer=TOKENIZER),
|
||||
structured_outputs_config=StructuredOutputsConfig(backend="guidance"),
|
||||
parallel_config=ParallelConfig(distributed_executor_backend=executor_backend),
|
||||
)
|
||||
|
||||
@@ -7,7 +7,7 @@ from unittest.mock import Mock
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.config import ModelConfig, RendererConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.reasoning import ReasoningParser
|
||||
from vllm.v1.request import Request
|
||||
from vllm.v1.structured_output import StructuredOutputManager
|
||||
@@ -17,19 +17,26 @@ class TestReasoningStructuredOutput:
|
||||
"""Test reasoning-aware structured output functionality."""
|
||||
|
||||
@pytest.fixture
|
||||
def mock_model_config(self):
|
||||
"""Create a mock ModelConfig."""
|
||||
config = Mock(spec=ModelConfig)
|
||||
config.skip_tokenizer_init = True # Skip tokenizer init to avoid network calls
|
||||
config.get_vocab_size = Mock(return_value=50000)
|
||||
def mock_renderer_config(self):
|
||||
"""Create a mock RendererConfig."""
|
||||
renderer_config = Mock(spec=RendererConfig)
|
||||
renderer_config.skip_tokenizer_init = (
|
||||
True # Skip tokenizer init to avoid network calls
|
||||
)
|
||||
|
||||
model_config = Mock(spec=ModelConfig)
|
||||
model_config.get_vocab_size = Mock(return_value=50000)
|
||||
model_config.trust_remote_code = False
|
||||
# Add missing runner_type attribute that tokenizer initialization expects
|
||||
config.runner_type = "generate"
|
||||
model_config.runner_type = "generate"
|
||||
renderer_config.model_config = model_config
|
||||
|
||||
# Add other attributes that tokenizer initialization might need
|
||||
config.tokenizer = "test-tokenizer"
|
||||
config.tokenizer_mode = "auto"
|
||||
config.trust_remote_code = False
|
||||
config.tokenizer_revision = None
|
||||
return config
|
||||
renderer_config.tokenizer = "test-tokenizer"
|
||||
renderer_config.tokenizer_mode = "auto"
|
||||
renderer_config.tokenizer_revision = None
|
||||
|
||||
return renderer_config
|
||||
|
||||
@pytest.fixture
|
||||
def mock_scheduler_config(self):
|
||||
@@ -39,10 +46,10 @@ class TestReasoningStructuredOutput:
|
||||
return config
|
||||
|
||||
@pytest.fixture
|
||||
def mock_vllm_config(self, mock_model_config, mock_scheduler_config):
|
||||
def mock_vllm_config(self, mock_renderer_config, mock_scheduler_config):
|
||||
"""Create a mock VllmConfig."""
|
||||
config = Mock(spec=VllmConfig)
|
||||
config.model_config = mock_model_config
|
||||
config.renderer_config = mock_renderer_config
|
||||
config.scheduler_config = mock_scheduler_config
|
||||
config.structured_outputs_config = Mock()
|
||||
config.structured_outputs_config.reasoning_parser = None
|
||||
|
||||
@@ -7,6 +7,7 @@ from vllm.attention.layer import Attention
|
||||
from vllm.config import (
|
||||
CacheConfig,
|
||||
ModelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
VllmConfig,
|
||||
set_current_vllm_config,
|
||||
@@ -45,6 +46,7 @@ def get_vllm_config():
|
||||
)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
scheduler_config=scheduler_config,
|
||||
)
|
||||
|
||||
@@ -13,6 +13,7 @@ from vllm.config import (
|
||||
CacheConfig,
|
||||
ModelConfig,
|
||||
ParallelConfig,
|
||||
RendererConfig,
|
||||
SchedulerConfig,
|
||||
VllmConfig,
|
||||
set_current_vllm_config,
|
||||
@@ -101,6 +102,7 @@ def get_vllm_config():
|
||||
parallel_config = ParallelConfig()
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
scheduler_config=scheduler_config,
|
||||
parallel_config=parallel_config,
|
||||
@@ -811,6 +813,7 @@ def test_hybrid_attention_mamba_tensor_shapes():
|
||||
attention_config = AttentionConfig(backend=AttentionBackendEnum.FLASHINFER)
|
||||
vllm_config = VllmConfig(
|
||||
model_config=model_config,
|
||||
renderer_config=RendererConfig(model_config=model_config),
|
||||
cache_config=cache_config,
|
||||
scheduler_config=scheduler_config,
|
||||
parallel_config=parallel_config,
|
||||
|
||||
Reference in New Issue
Block a user