241 lines
9.8 KiB
Python
241 lines
9.8 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""
|
|
Diffusion Model Runner for vLLM-Omni.
|
|
|
|
Handles model loading, compilation, caching, and execution of diffusion model
|
|
forward passes. This follows the AR pattern where the Runner handles all
|
|
model-related operations.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import time
|
|
from collections.abc import Iterable
|
|
from contextlib import nullcontext
|
|
|
|
import torch
|
|
from torch.profiler import record_function
|
|
from vllm.config import LoadConfig
|
|
from vllm.logger import init_logger
|
|
from vllm.utils.mem_utils import DeviceMemoryProfiler, GiB_bytes
|
|
|
|
from vllm_omni.diffusion.cache.cache_dit_backend import cache_summary
|
|
from vllm_omni.diffusion.cache.selector import get_cache_backend
|
|
from vllm_omni.diffusion.compile import regionally_compile
|
|
from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
|
|
from vllm_omni.diffusion.forward_context import set_forward_context
|
|
from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
|
|
from vllm_omni.diffusion.offloader import get_offload_backend
|
|
from vllm_omni.diffusion.registry import _NO_CACHE_ACCELERATION
|
|
from vllm_omni.diffusion.request import OmniDiffusionRequest
|
|
from vllm_omni.distributed.omni_connectors.kv_transfer_manager import OmniKVTransferManager
|
|
from vllm_omni.platforms import current_omni_platform
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
class DiffusionModelRunner:
|
|
"""
|
|
Model runner that handles model loading and execution for diffusion models.
|
|
|
|
This class follows the AR pattern where the Runner handles all model-related
|
|
operations including loading, compilation, offloading, caching, and execution.
|
|
The Worker only handles infrastructure (device, distributed env).
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
vllm_config,
|
|
od_config: OmniDiffusionConfig,
|
|
device: torch.device,
|
|
):
|
|
"""
|
|
Initialize the diffusion model runner.
|
|
|
|
Args:
|
|
vllm_config: vLLM configuration.
|
|
od_config: OmniDiffusion configuration.
|
|
device: The device to run on.
|
|
"""
|
|
self.vllm_config = vllm_config
|
|
self.od_config = od_config
|
|
self.device = device
|
|
self.pipeline = None
|
|
self.cache_backend = None
|
|
self.offload_backend = None
|
|
|
|
# Initialize KV cache manager for connector management
|
|
self.kv_transfer_manager = OmniKVTransferManager.from_od_config(od_config)
|
|
|
|
def _compile_transformer(self, attr_name: str) -> None:
|
|
"""Compile a transformer attribute on the pipeline with torch.compile."""
|
|
model = getattr(self.pipeline, attr_name, None)
|
|
if model is None:
|
|
return
|
|
try:
|
|
setattr(self.pipeline, attr_name, regionally_compile(model, dynamic=True))
|
|
logger.info("Model runner: %s compiled with torch.compile.", attr_name)
|
|
except Exception as e:
|
|
logger.warning(
|
|
"Model runner: torch.compile for %s failed: %s. Using eager mode.",
|
|
attr_name,
|
|
e,
|
|
)
|
|
|
|
def load_model(
|
|
self,
|
|
memory_pool_context_fn: callable | None = None,
|
|
load_format: str | None = None,
|
|
custom_pipeline_name: str | None = None,
|
|
) -> None:
|
|
"""
|
|
Load the diffusion model, apply compilation and offloading.
|
|
|
|
Args:
|
|
memory_pool_context_fn: Optional function that returns a context manager
|
|
for memory pool allocation (used for sleep mode).
|
|
load_format: Format for loading model weights. Supported formats:
|
|
- "default" (default): Automatically detect and use the default format based on configuration
|
|
- "custom_pipeline": Init model from a custom pipeline class specified by `custom_pipeline_name`
|
|
- "dummy": Skip actual weight loading, useful for testing and custom pipelines that
|
|
don't require default weights.
|
|
custom_pipeline_name: Optional custom pipeline class name to use.
|
|
"""
|
|
|
|
if load_format == "dummy":
|
|
return
|
|
|
|
load_device = (
|
|
"cpu" if self.od_config.enable_cpu_offload or self.od_config.enable_layerwise_offload else str(self.device)
|
|
)
|
|
|
|
def get_memory_context():
|
|
if memory_pool_context_fn is not None:
|
|
return memory_pool_context_fn(tag="weights")
|
|
return nullcontext()
|
|
|
|
# Load model within forward context
|
|
load_config = LoadConfig()
|
|
model_loader = DiffusersPipelineLoader(load_config, od_config=self.od_config)
|
|
time_before_load = time.perf_counter()
|
|
|
|
with get_memory_context():
|
|
with DeviceMemoryProfiler() as m:
|
|
self.pipeline = model_loader.load_model(
|
|
od_config=self.od_config,
|
|
load_device=load_device,
|
|
load_format=load_format,
|
|
custom_pipeline_name=custom_pipeline_name,
|
|
device=self.device,
|
|
)
|
|
time_after_load = time.perf_counter()
|
|
|
|
logger.info(
|
|
"Model loading took %.4f GiB and %.6f seconds",
|
|
m.consumed_memory / GiB_bytes,
|
|
time_after_load - time_before_load,
|
|
)
|
|
logger.info("Model runner: Model loaded successfully.")
|
|
|
|
# Apply CPU offloading
|
|
self.offload_backend = get_offload_backend(self.od_config, device=self.device)
|
|
if self.offload_backend is not None:
|
|
logger.info(f" Enabling offloader backend: {self.offload_backend.__class__.__name__}")
|
|
self.offload_backend.enable(self.pipeline)
|
|
|
|
# Apply torch.compile if not in eager mode
|
|
if not self.od_config.enforce_eager:
|
|
if current_omni_platform.supports_torch_inductor():
|
|
self._compile_transformer("transformer")
|
|
self._compile_transformer("transformer_2")
|
|
else:
|
|
logger.warning(
|
|
"Model runner: Platform %s does not support torch inductor, skipping torch.compile.",
|
|
current_omni_platform.get_torch_device(),
|
|
)
|
|
|
|
# Setup cache backend
|
|
self.cache_backend = get_cache_backend(self.od_config.cache_backend, self.od_config.cache_config)
|
|
|
|
if self.cache_backend is not None:
|
|
if self.od_config.model_class_name in _NO_CACHE_ACCELERATION:
|
|
logger.warning(
|
|
"Cache backend '%s' is not supported for %s; disabling cache acceleration.",
|
|
self.od_config.cache_backend,
|
|
self.od_config.model_class_name,
|
|
)
|
|
self.cache_backend = None
|
|
self.od_config.cache_backend = None
|
|
else:
|
|
self.cache_backend.enable(self.pipeline)
|
|
|
|
logger.info("Model runner: Initialization complete.")
|
|
|
|
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
|
"""Load weights into the pipeline."""
|
|
return self.pipeline.load_weights(weights)
|
|
|
|
def execute_model(self, req: OmniDiffusionRequest) -> DiffusionOutput:
|
|
"""
|
|
Execute a forward pass for the given requests.
|
|
|
|
Args:
|
|
req: A diffusion request containing a list of prompts to process.
|
|
|
|
Returns:
|
|
DiffusionOutput with generated results.
|
|
|
|
Note:
|
|
We use torch.no_grad() for HSDP because HSDP2's fully_shard requires access
|
|
to tensor version counters in pre_forward hooks, which inference tensors do
|
|
not track. For non-HSDP inference, we use torch.inference_mode() for better
|
|
performance.
|
|
"""
|
|
assert self.pipeline is not None, "Model not loaded. Call load_model() first."
|
|
if len(req.prompts) == 0:
|
|
raise ValueError("Cannot execute model with empty request list")
|
|
|
|
# Use no_grad() for HSDP compatibility, inference_mode() otherwise for better perf
|
|
use_hsdp = self.od_config.parallel_config.use_hsdp
|
|
grad_context = torch.no_grad() if use_hsdp else torch.inference_mode()
|
|
with grad_context:
|
|
# The manager handles the check for need_recv_cache internally
|
|
self.kv_transfer_manager.receive_multi_kv_cache(
|
|
req,
|
|
cfg_kv_collect_func=getattr(self.od_config, "cfg_kv_collect_func", None),
|
|
target_device=getattr(self.pipeline, "device", None),
|
|
)
|
|
|
|
if req.sampling_params.generator is None and req.sampling_params.seed is not None:
|
|
if req.sampling_params.generator_device is not None:
|
|
gen_device = req.sampling_params.generator_device
|
|
elif self.device.type == "cpu":
|
|
gen_device = "cpu"
|
|
else:
|
|
gen_device = self.device
|
|
req.sampling_params.generator = torch.Generator(device=gen_device).manual_seed(req.sampling_params.seed)
|
|
|
|
# Refresh cache context if needed
|
|
if (
|
|
not getattr(req, "skip_cache_refresh", False)
|
|
and self.cache_backend is not None
|
|
and self.cache_backend.is_enabled()
|
|
):
|
|
self.cache_backend.refresh(self.pipeline, req.sampling_params.num_inference_steps)
|
|
|
|
with set_forward_context(vllm_config=self.vllm_config, omni_diffusion_config=self.od_config):
|
|
with record_function("pipeline_forward"):
|
|
output = self.pipeline.forward(req)
|
|
|
|
# NOTE:
|
|
if (
|
|
self.cache_backend is not None
|
|
and self.cache_backend.is_enabled()
|
|
and self.od_config.cache_backend == "cache_dit"
|
|
and self.od_config.enable_cache_dit_summary
|
|
):
|
|
cache_summary(self.pipeline, details=True)
|
|
|
|
return output
|