[Core] Unified quantization framework (#1764)
This commit is contained in:
1 parent
2a630313ce
commit
6b93459826
40 files changed
+3306
-605
No files matched your search
+25
-37
@@ -13,18 +13,16 @@ from pydantic import model_validator
|
||||
from typing_extensions import Self
|
||||
from vllm.config.utils import config
|
||||
from vllm.logger import init_logger
|
||||
|
||||
from vllm_omni.diffusion.quantization import (
|
||||
DiffusionQuantizationConfig,
|
||||
get_diffusion_quant_config,
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig,
|
||||
)
|
||||
|
||||
from vllm_omni.diffusion.utils.network_utils import is_port_available
|
||||
from vllm_omni.quantization import build_quant_config
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.config import ProfilerConfig
|
||||
|
||||
from vllm_omni.diffusion.quantization import DiffusionQuantizationConfig
|
||||
|
||||
# Import after TYPE_CHECKING to avoid circular imports at runtime
|
||||
# The actual import is deferred to __post_init__ to avoid import order issues
|
||||
|
||||
@@ -483,10 +481,10 @@ class OmniDiffusionConfig:
|
||||
# Model-specific function for collecting CFG KV caches (set at runtime)
|
||||
cfg_kv_collect_func: Any | None = None
|
||||
|
||||
# Quantization settings
|
||||
# Supported methods: "fp8" (FP8 W8A8 on Ada/Hopper, weight-only on older GPUs)
|
||||
quantization: str | None = None
|
||||
quantization_config: "DiffusionQuantizationConfig | dict[str, Any] | None" = None
|
||||
# Quantization: str method name, dict config, QuantizationConfig, or None.
|
||||
# str is resolved to {"method": <str>} internally.
|
||||
# Per-component: {"transformer": {"method": "fp8"}, "vae": None}
|
||||
quantization_config: str | QuantizationConfig | dict[str, Any] | None = None
|
||||
|
||||
# Diffusion pipeline Profiling config
|
||||
enable_diffusion_pipeline_profiler: bool = False
|
||||
@@ -597,34 +595,17 @@ class OmniDiffusionConfig:
|
||||
# If it's neither dict nor DiffusionCacheConfig, convert to empty config
|
||||
self.cache_config = DiffusionCacheConfig()
|
||||
|
||||
# Convert quantization config (deferred import to avoid circular imports)
|
||||
if self.quantization is not None or self.quantization_config is not None:
|
||||
from vllm_omni.diffusion.quantization import (
|
||||
DiffusionQuantizationConfig,
|
||||
)
|
||||
|
||||
# Handle dict or DictConfig (from OmegaConf) - use Mapping for broader compatibility
|
||||
if isinstance(self.quantization_config, Mapping):
|
||||
# Convert DictConfig to dict if needed (OmegaConf compatibility)
|
||||
config_dict = dict(self.quantization_config)
|
||||
# Use get() instead of pop() to avoid mutating original dict
|
||||
quant_method = config_dict.get("method", self.quantization)
|
||||
# Filter out "method" key for kwargs
|
||||
quant_kwargs = {k: v for k, v in config_dict.items() if k != "method"}
|
||||
|
||||
# Validate conflicting methods
|
||||
if self.quantization is not None and quant_method is not None and quant_method != self.quantization:
|
||||
logger.warning(
|
||||
f"Conflicting quantization methods: quantization={self.quantization!r}, "
|
||||
f"quantization_config['method']={quant_method!r}. Using quantization_config['method']."
|
||||
)
|
||||
|
||||
self.quantization_config = get_diffusion_quant_config(quant_method, **quant_kwargs)
|
||||
elif self.quantization_config is None and self.quantization is not None:
|
||||
self.quantization_config = get_diffusion_quant_config(self.quantization)
|
||||
elif not isinstance(self.quantization_config, DiffusionQuantizationConfig):
|
||||
# Resolve quantization_config: str/dict -> QuantizationConfig via build_quant_config.
|
||||
if self.quantization_config is not None:
|
||||
if isinstance(self.quantization_config, QuantizationConfig):
|
||||
pass # Already built
|
||||
elif isinstance(self.quantization_config, str):
|
||||
self.quantization_config = build_quant_config(self.quantization_config)
|
||||
elif isinstance(self.quantization_config, Mapping):
|
||||
self.quantization_config = build_quant_config(dict(self.quantization_config))
|
||||
else:
|
||||
raise TypeError(
|
||||
f"quantization_config must be a DiffusionQuantizationConfig, dict, or None, "
|
||||
f"quantization_config must be str, dict, QuantizationConfig, or None, "
|
||||
f"got {type(self.quantization_config)!r}"
|
||||
)
|
||||
|
||||
@@ -646,6 +627,13 @@ class OmniDiffusionConfig:
|
||||
kwargs["lora_scale"] = kwargs["static_lora_scale"]
|
||||
kwargs.pop("static_lora_scale", None)
|
||||
|
||||
# Backwards-compatibility: map "quantization" to "quantization_config"
|
||||
# so callers using the old field name still work.
|
||||
if "quantization" in kwargs and "quantization_config" not in kwargs:
|
||||
kwargs["quantization_config"] = kwargs.pop("quantization")
|
||||
else:
|
||||
kwargs.pop("quantization", None)
|
||||
|
||||
# Check environment variable as fallback for cache_backend
|
||||
# Support both old DIFFUSION_CACHE_ADAPTER and new DIFFUSION_CACHE_BACKEND for backwards compatibility
|
||||
if "cache_backend" not in kwargs:
|
||||
|
||||
Reference in new issue
Block a user