update ar_dit_test
This commit is contained in:
1 parent
ff9c7d4536
commit
81f04d3cf9
32 files changed
+1192
-1908
No files matched your search
@@ -2,21 +2,22 @@
|
||||
|
||||
## Overview
|
||||
|
||||
vLLM-omni is a multi-modality extension for vLLM that supports non-autoregressive structures and non-textual outputs. This document outlines the key software abstractions, APIs, and dependencies for the system, designed to maximize reuse of vLLM's proven architecture.
|
||||
vLLM-omni extends vLLM with multi-stage, multimodal capabilities. It supports both autoregressive (AR) stages and diffusion (DiT) stages, enabling pipelines such as text→image generation.
|
||||
|
||||
## Architecture Principles
|
||||
|
||||
1. **vLLM V1 Compatibility**: Built on vLLM's Engine V1 architecture with AsyncLLM and EngineCore patterns
|
||||
2. **Stage-based Processing**: Models are divided into multiple stages, each processed by different Engine Cores
|
||||
3. **Multiple Engine Core Support**: Each stage can use either AR Engine Core (reusing vLLM) or Diffusion Engine Core (new DiT support), or other new Cores
|
||||
4. **Worker Process Pattern**: Follows vLLM's multiprocess worker architecture for scalability
|
||||
5. **Extensibility**: Easy integration of new modalities, model architectures, and output formats
|
||||
1. **vLLM V1 Compatibility** – Reuse LLMEngine/AsyncLLM and executor abstractions.
|
||||
2. **Stage-based Processing** – Each stage encapsulates its own engine and configuration.
|
||||
3. **Executor/Worker Separation** – Executors remain thin; model-specific logic lives in workers and runners.
|
||||
4. **Extensibility** – New modalities can be added by implementing dedicated stage configs, workers, and output processors.
|
||||
|
||||
## Key Data Flow
|
||||
|
||||
API Server --> OmniLLM/AsyncOmniLLM (New, including multi engines) --> LLMEngine/AsyncLLM --> Engine Core
|
||||
--> Scheduler (New one for DiT) --> Executor (New one for diffusers) --> Worker (New one for DiT)
|
||||
--> ModelRunner (New one for AR hiddenstate, New one for DiT) --> RequestState --> OutputProcessor (New one for final multimodal output)
|
||||
```
|
||||
Request → OmniLLM/AsyncOmniLLM → StageManager → LLMEngine/Executor → Worker → ModelRunner → OutputProcessor → Response
|
||||
```
|
||||
|
||||
For diffusion stages, the executor hands off to `DiffusionGPUWorker`, which wraps `DiffusionModelRunner` (responsible for diffusers pipeline loading and inference).
|
||||
|
||||
## Core Components (Ordered by Data Flow)
|
||||
|
||||
@@ -27,399 +28,130 @@ pip install vllm-omni
|
||||
```
|
||||
|
||||
### 2. Online inference launch (Entry Point)
|
||||
|
||||
Keep consistency with the vllm main branch
|
||||
```bash
|
||||
vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8000
|
||||
vllm serve Qwen/Qwen3-0.6B --omni --port 8000
|
||||
```
|
||||
|
||||
#### Design Architecture
|
||||
```mermaid
|
||||
graph TD
|
||||
A[vllm serve model --omni] --> B[vLLM-omni CLI Wrapper]
|
||||
B --> C{Detect --omni flag}
|
||||
C -->|Yes| D[Parse omniConfig]
|
||||
C -->|No| E[Forward to vLLM CLI]
|
||||
D --> F[Initialize AsyncOmniLLM]
|
||||
F --> G[Start Omni Server]
|
||||
G --> H[Multi-stage Processing]
|
||||
E --> I[Standard vLLM Pipeline]
|
||||
|
||||
subgraph "vLLM-omni Components"
|
||||
F
|
||||
G
|
||||
H
|
||||
end
|
||||
|
||||
subgraph "vLLM Components"
|
||||
I
|
||||
end
|
||||
```
|
||||
|
||||
#### Omni Serve Command Implementation
|
||||
File: vllm_omni/entrypoints/cli/main.py
|
||||
### 3. Omni Serve Command (CLI)
|
||||
```python
|
||||
import sys
|
||||
import argparse
|
||||
from typing import List, Optional
|
||||
from vllm_omni.entrypoints.omni import OmniServeCommand
|
||||
|
||||
def main():
|
||||
"""Main CLI entry point that intercepts vLLM commands"""
|
||||
# Check if --omni flag is present
|
||||
if "--omni" in sys.argv:
|
||||
# Remove --omni flag and process with vLLM-omni
|
||||
omni_args = [arg for arg in sys.argv[1:] if arg != "--omni"]
|
||||
omni_serve = OmniServeCommand()
|
||||
omni_serve.run(omni_args)
|
||||
else:
|
||||
# Forward to original vLLM CLI
|
||||
from vllm.entrypoints.cli.main import main as vllm_main
|
||||
vllm_main()
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
# vllm_omni/entrypoints/cli/main.py
|
||||
if "--omni" in sys.argv:
|
||||
omni_args = [arg for arg in sys.argv[1:] if arg not in {"--omni", "serve"}]
|
||||
OmniServeCommand().run(omni_args)
|
||||
else:
|
||||
from vllm.entrypoints.cli.main import main as vllm_main
|
||||
vllm_main()
|
||||
```
|
||||
|
||||
#### vLLM Plugin Integration
|
||||
File: vllm_omni/pyproject.toml
|
||||
### 4. Offline Inference (OmniLLM)
|
||||
```python
|
||||
[project.scripts]
|
||||
# Override vLLM's CLI entry point
|
||||
vllm = "vllm_omni.cli:main"
|
||||
|
||||
# Add entry point for vLLM integration
|
||||
[project.entry-points."vllm.plugins"]
|
||||
omni = "vllm_omni.plugin:OmniPlugin"
|
||||
```
|
||||
|
||||
### 3. Offline inference launch (Entry Point)
|
||||
Design an upper level class to incorporate multi Engines, each engine has a engine call.
|
||||
```python
|
||||
from vllm.entrypoints.llm import LLM
|
||||
from vllm.v1.engine.llm_engine import LLMEngine
|
||||
|
||||
# vllm_omni/entrypoints/omni_llm.py
|
||||
class OmniLLM(LLM):
|
||||
"""Extended LLM supporting multiple engines and stage-based processing"""
|
||||
|
||||
def __init__(self, stage_configs: List[OmniStageConfig]):
|
||||
super().__init__()
|
||||
self.stage_configs = stage_configs
|
||||
self.engine_list = [] # List of AsyncLLM instances for each stage
|
||||
def __init__(..., stage_configs):
|
||||
super().__init__(model=stage_configs[0].model_path if stage_configs else "test-model")
|
||||
self.stage_manager = StageManager(stage_configs)
|
||||
self.output_processor = MultimodalOutputProcessor()
|
||||
self._initialize_stage_engines()
|
||||
|
||||
def _initialize_stage_engines(self) -> None:
|
||||
"""Initialize LLMEngine instances for each stage"""
|
||||
for stage_config in self.stage_configs:
|
||||
stage_llm = LLMEngine.from_vllm_config(
|
||||
vllm_config=stage_config.vllm_config,
|
||||
executor_class=self.executor_class,
|
||||
log_stats=self.log_stats
|
||||
)
|
||||
self.engine_list.append(stage_llm)
|
||||
|
||||
def generate(
|
||||
self,
|
||||
stage_args_list: List[stage_args],
|
||||
use_tqdm: Union[bool, Callable[..., tqdm]] = True,
|
||||
lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
|
||||
priority: Optional[list[int]] = None,
|
||||
) -> list[RequestOutput]:
|
||||
"""Main generation interface - orchestrates multi-stage processing"""
|
||||
self.stage_manager.initialize_engines()
|
||||
|
||||
# Process through each stage sequentially
|
||||
for i, stage_config in enumerate(self.stage_configs):
|
||||
stage_engine = self.engine_list[i]
|
||||
|
||||
# Prepare input for this stage
|
||||
stage_args = stage_args_list[i]
|
||||
prompt_str, engine_request, tokenization_kwargs = self._process_stage_inputs(stage_config, **stage_args)
|
||||
|
||||
# Add inputs to Engine
|
||||
stage_engine.add_request(requesy_id, prompt_str, tokenization_kwargs)
|
||||
# Run Engine
|
||||
stage_output = stage_engine.step()
|
||||
|
||||
# Update stage input and output of each stage for later usage
|
||||
self._update_stage_io(stage_output, stage_config)
|
||||
|
||||
|
||||
# Process final output
|
||||
final_output = self.output_processor.process_output(
|
||||
self.stage_configs[-1].stage_output
|
||||
)
|
||||
return final_output
|
||||
|
||||
def _process_stage_inputs(stage_config, **stage_args) -> Any:
|
||||
"""Prepare input for specific stage"""
|
||||
if stage_config.engine_type == "AR":
|
||||
return self._process_ar_inputs(**stage_args)
|
||||
elif stage_config.engine_type == "DiT":
|
||||
return self._process_dit_inputs(**stage_args)
|
||||
else:
|
||||
raise NotImplementedError
|
||||
|
||||
def _process_dit_inputs(**stage_args)-> Any:
|
||||
image_latent = self.vae.encode(**stage_args)
|
||||
|
||||
return image_latent
|
||||
def generate(...):
|
||||
current_output = None
|
||||
for idx, (config, args) in enumerate(zip(self.stage_configs, stage_args_list)):
|
||||
engine = self.stage_manager.get_engine(idx)
|
||||
processed = self._process_stage_inputs(config, args, current_output)
|
||||
stage_output = self._execute_stage(engine, processed, stage_config=config)
|
||||
current_output = stage_output
|
||||
config.stage_output = stage_output
|
||||
return self.output_processor.process_output(current_output)
|
||||
```
|
||||
|
||||
### 4. Stage config for setting of different model stages
|
||||
### 5. Stage Manager
|
||||
```python
|
||||
# vllm_omni/core/stage_manager.py
|
||||
class StageManager:
|
||||
def initialize_engines(self):
|
||||
if self._initialized:
|
||||
return
|
||||
for config in self.stage_configs:
|
||||
if config.engine_type == "AR":
|
||||
engine = self._create_ar_engine(config)
|
||||
elif config.engine_type == "DiT":
|
||||
engine = self._create_dit_engine(config)
|
||||
self.engine_list.append(engine)
|
||||
self._initialized = True
|
||||
```
|
||||
|
||||
### 6. Diffusion Execution (DiT Stage)
|
||||
```python
|
||||
# vllm_omni/worker/diffusion_model_runner.py
|
||||
class DiffusionModelRunner:
|
||||
def __init__(self, model_path, pipeline_name=None, device=None, dtype=None):
|
||||
self.device, self.torch_dtype = self._resolve_device_and_dtype(device, dtype)
|
||||
self._pipeline = self._load_pipeline(model_path, pipeline_name)
|
||||
|
||||
def generate(...):
|
||||
output = self._pipeline(...)
|
||||
return DiffusionRunnerOutput(prompt=prompt, images=getattr(output, "images", []) or [])
|
||||
```
|
||||
|
||||
```python
|
||||
# vllm_omni/worker/gpu_diffusion_worker.py
|
||||
class DiffusionGPUWorker:
|
||||
def __init__(self, model_path, ...):
|
||||
self.model_runner = DiffusionModelRunner(model_path, ...)
|
||||
|
||||
def generate(...):
|
||||
return self.model_runner.generate(...)
|
||||
```
|
||||
|
||||
```python
|
||||
# vllm_omni/executor/diffusers_executor.py
|
||||
class DiffusersPipelineExecutor:
|
||||
def __init__(self, model_path, ...):
|
||||
self.worker = DiffusionGPUWorker(model_path, ...)
|
||||
|
||||
def generate(...):
|
||||
return self.worker.generate(...)
|
||||
```
|
||||
|
||||
This mirrors the standard executor→worker→runner layering in vLLM.
|
||||
|
||||
### 7. Stage Configuration
|
||||
```python
|
||||
# vllm_omni/config/stage_config.py
|
||||
@dataclass
|
||||
class OmniStageConfig:
|
||||
"""Configuration for a processing stage"""
|
||||
stage_id: int
|
||||
engine_type: str # "AR" or "DiT"
|
||||
engine_type: Literal["AR", "DiT"]
|
||||
model_path: str
|
||||
input_modalities: List[str]
|
||||
output_modalities: List[str]
|
||||
vllm_config: Optional[VllmConfig] = None # For engine config of corresponding stage
|
||||
executor_class: type[Executor] # For execute class config of corresponding stage
|
||||
dit_config: Optional[DiTConfig] = None # For diffusion stages
|
||||
cache_config: Optional[DiTCacheConfig] = None
|
||||
vllm_config: Optional[VllmConfig] = None
|
||||
executor_class: Optional[type[Executor]] = None
|
||||
dit_config: Optional[DiTConfig] = None
|
||||
```
|
||||
For AR stage, the setting is:
|
||||
```python
|
||||
from vllm.v1.core.sched.scheduler import Scheduler
|
||||
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
|
||||
ar_stage_config = OmniStageConfig()
|
||||
ar_stage_config.vllm_config.scheduler_config.scheduler_cls = Scheduler # original vllm scheduler for AR
|
||||
ar_stage_config.executor_class = MultiprocExecutor # original vllm executor for AR
|
||||
```
|
||||
For DiT stage, the setting is:
|
||||
```python
|
||||
from vllm.v1.executor.multiproc_executor import MultiprocExecutor
|
||||
dit_stage_config = OmniStageConfig()
|
||||
dit_stage_config.vllm_config.scheduler_config.scheduler_cls = DiffusionScheduler # New scheduler for DiT
|
||||
|
||||
# For diffusion models without using diffusers
|
||||
dit_stage_config.executor_class = MultiprocExecutor
|
||||
- AR stages default to `UniProcExecutor`.
|
||||
- DiT stages set `executor_class` to `DiffusersPipelineExecutor` when `use_diffusers=True`.
|
||||
|
||||
# For diffusion models using diffusers
|
||||
dit_stage_config.executor_class = DiffusersPipelineExecutor
|
||||
```
|
||||
|
||||
### 5. Online inference main class
|
||||
|
||||
Similar to OmniLLM in offline inference, add some asynchronous processing, referring to AsyncLLM
|
||||
```python
|
||||
from vllm.v1.engine.async_llm import AsyncLLM
|
||||
|
||||
|
||||
class AsyncOmniLLM(AsyncLLM):
|
||||
"""Extended AsyncLLM supporting multiple engines and stage-based processing"""
|
||||
|
||||
def __init__(self, stage_configs: List[OmniStageConfig]):
|
||||
super().__init__()
|
||||
```
|
||||
|
||||
### 6. Engine Core
|
||||
No need to change. The specific executor, scheduler and output type can be transferred into it with new configs.
|
||||
```python
|
||||
# class: EngineCore
|
||||
# EngineCore.step (simplified)
|
||||
class EngineCore:
|
||||
def step(self):
|
||||
scheduler_output = scheduler.schedule()
|
||||
model_output = executor.execute_model(scheduler_output)
|
||||
engine_outputs = scheduler.update_from_output(
|
||||
scheduler_output, model_output
|
||||
)
|
||||
return engine_outputs
|
||||
```
|
||||
|
||||
### 7. Scheduler
|
||||
Create a new Diffusion Scheduler for DiT, which is inherited from original vllm AR scheduler. At first, no complicated strategy.
|
||||
Just process the request first in first out. Then create a class of DiT Cache Manager for optimization in the future.
|
||||
```python
|
||||
from vllm.v1.core.sched.scheduler import SchedulerInterface
|
||||
|
||||
|
||||
class OmniDiffusionScheduler(SchedulerInterface):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
kv_cache_config: KVCacheConfig,
|
||||
dit_cache_config: DiTCacheConfig,
|
||||
structured_output_manager: StructuredOutputManager,
|
||||
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
|
||||
include_finished_set: bool = False,
|
||||
log_stats: bool = False,
|
||||
) -> None:
|
||||
|
||||
self.dit_cache_manager = DiTCacheManager(dit_cache_config)
|
||||
```
|
||||
For DiT Cache Manager (Can refer to xDiT):
|
||||
```python
|
||||
class DiTCacheManager:
|
||||
"""Manages DiT-specific caching"""
|
||||
|
||||
def __init__(self, config: DiTCacheConfig):
|
||||
self.cache_tensors: Dict[str, torch.Tensor] = {}
|
||||
self.cache_groups: List[DiTCacheTensor] = config.dit_cache_tensors
|
||||
|
||||
def allocate_cache(self, request_id: str, size: int) -> torch.Tensor
|
||||
def get_cache(self, request_id: str) -> Optional[torch.Tensor]
|
||||
def release_cache(self, request_id: str) -> None
|
||||
def clear_expired_cache(self) -> None
|
||||
```
|
||||
|
||||
### 8. Executor for DiT without diffusers
|
||||
|
||||
No need to change. Just import original AR executor.
|
||||
|
||||
|
||||
|
||||
### 9. Worker
|
||||
|
||||
#### Inheritance strategy
|
||||
Prefer reusing the mature GPU Worker end-to-end. Worker class is selected by configuration (vllm_config). Do not add a new executor-specific worker binding. If customization is needed, override only `init_device` to construct the `OmniDiffusionModelRunner`; all other behaviors (device init details, profiling, sleep/wake, PP/TP comms, execute path) remain from `vllm/v1/worker/gpu_worker.py::Worker`.
|
||||
|
||||
|
||||
Inherited and overridden
|
||||
```python
|
||||
# Optional: only if you need to plug a custom OmniDiffusionModelRunner.
|
||||
from vllm.v1.worker.gpu_worker import Worker as GPUWorker
|
||||
from vllm.v1.worker.diffusion_model_runner import OmniDiffusionModelRunner
|
||||
|
||||
class Worker(GPUWorker):
|
||||
def init_device(self) -> None:
|
||||
#those related to device check and init
|
||||
...
|
||||
|
||||
self.model_runner = OmniDiffusionModelRunner(self.vllm_config, ...)
|
||||
```
|
||||
|
||||
### 10. Model Runner
|
||||
|
||||
### Function map (Model Runner)
|
||||
#### 1) Inherited and overridden
|
||||
Those parts relied to the KV Cache will be omitted if we do not register the model to the vllm config. The engine core will view it as do not require KV Cache, and handle it properly
|
||||
|
||||
Reuse `vllm/v1/outputs.py::ModelRunnerOutput`:
|
||||
- OmniDiffusionModelRunner: Use the `pooler_output=[Tensor,...]` to return multi modal tensors
|
||||
- OmniARModelRunner: Use the `pooler_output=[Tensor,...]` to return hidden states.
|
||||
```python
|
||||
from typing import Optional, Union
|
||||
import torch
|
||||
from vllm.v1.worker.gpu_model_runner import GPUModelRunner
|
||||
from vllm.v1.outputs import ModelRunnerOutput
|
||||
|
||||
|
||||
class OmniDiffusionModelRunner(GPUModelRunner):
|
||||
@torch.inference_mode()
|
||||
def execute_model(
|
||||
self,
|
||||
scheduler_output: "SchedulerOutput",
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
) -> Union[ModelRunnerOutput, IntermediateTensors]:
|
||||
...
|
||||
return ModelRunnerOutput(
|
||||
req_ids=[...],
|
||||
req_id_to_index={...},
|
||||
sampled_token_ids=[],
|
||||
spec_token_ids=None,
|
||||
logprobs=None,
|
||||
prompt_logprobs_dict={},
|
||||
pooler_output=[Tensor,...], # Return Hidden states
|
||||
kv_connector_output=None,
|
||||
num_nans_in_logits=None,
|
||||
)# return multi modal tensors via pooler_output=[Tensor,...]
|
||||
|
||||
|
||||
class OmniARModelRunner(GPUModelRunner):
|
||||
@torch.inference_mode()
|
||||
def execute_model(
|
||||
self,
|
||||
scheduler_output: "SchedulerOutput",
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
) -> Union[ModelRunnerOutput, IntermediateTensors]:
|
||||
...
|
||||
return ModelRunnerOutput(
|
||||
req_ids=[...],
|
||||
req_id_to_index={...},
|
||||
sampled_token_ids=[],
|
||||
spec_token_ids=None,
|
||||
logprobs=None,
|
||||
prompt_logprobs_dict={},
|
||||
pooler_output=[Tensor,...], # Return Hidden states
|
||||
kv_connector_output=None,
|
||||
num_nans_in_logits=None,
|
||||
)# return hidden states via pooler_output=[Tensor,...]
|
||||
```
|
||||
|
||||
### 11. RequestState Role and Usage
|
||||
```bash
|
||||
# RequestState serves as the per-request state tracker in OutputProcessor:
|
||||
# - Maintains request-specific state (tokens, logprobs, detokenizer, etc.)
|
||||
# - Converts EngineCoreOutput → RequestOutput/PoolingRequestOutput
|
||||
# - Manages request lifecycle from registration to completion
|
||||
AsyncLLM.add_request()
|
||||
↓
|
||||
OutputProcessor.add_request()
|
||||
↓
|
||||
RequestState.from_new_request() → 创建请求状态
|
||||
|
||||
AsyncLLM.__init__() / AsyncLLM.generate() / AsyncLLM.encode() → 创建一个Background loop 持续从EngineCore获取输出
|
||||
↓
|
||||
OutputProcessor.process_outputs() → 更新状态并处理输出
|
||||
↓
|
||||
RequestState.make_request_output() → 转换为最终输出,格式为RequestOutput或者PoolingRequestOutput
|
||||
↓
|
||||
RequestOutputCollector.put() → 推送到队列(AsyncLLM)
|
||||
```
|
||||
|
||||
Need to add implementation for an existing method
|
||||
```python
|
||||
class RequestState:
|
||||
def __init__(self, request_id: str, parent_req: Optional[ParentRequest],
|
||||
request_index: int, lora_name: Optional[str],
|
||||
output_kind: RequestOutputKind, prompt: Optional[str],
|
||||
prompt_token_ids: list[int], logprobs_processor: Optional[LogprobsProcessor],
|
||||
detokenizer: Optional[IncrementalDetokenizer],
|
||||
max_tokens_param: Optional[int], arrival_time: float,
|
||||
queue: Optional[RequestOutputCollector], log_stats: bool):
|
||||
|
||||
def _new_pooling_output(self, pooling_output: torch.Tensor) -> PoolingOutput:
|
||||
"""Create PoolingOutput for multimodal/pooling requests"""
|
||||
return PoolingOutput(data=pooling_output)
|
||||
```
|
||||
|
||||
|
||||
### 12. OutputProcessor
|
||||
|
||||
For hidden state output, original OutputProcessor already support.
|
||||
|
||||
We only need to add new one for final multimodal output.
|
||||
### 8. Output Processing
|
||||
```python
|
||||
# vllm_omni/engine/output_processor.py
|
||||
class MultimodalOutputProcessor(OutputProcessor):
|
||||
"""Handles multimodal output processing"""
|
||||
|
||||
def __init__(self):
|
||||
self.output_handlers: Dict[str, OutputProcessor] = {}
|
||||
|
||||
def process_outputs(self, engine_core_outputs: list[EngineCoreOutput], ...):
|
||||
for engine_core_output in engine_core_outputs:
|
||||
# Option 1: Use output_type field (if EngineCoreOutput is extended)
|
||||
if engine_core_output.output_type == "image":
|
||||
self._process_image_output(engine_core_output)
|
||||
elif engine_core_output.output_type == "text+image":
|
||||
self._process_text_image_output(engine_core_output)
|
||||
elif engine_core_output.output_type == "latents":
|
||||
self._process_latents_output(engine_core_output)
|
||||
elif engine_core_output.output_type == "text":
|
||||
self._process_text_output(engine_core_output)
|
||||
else:
|
||||
# Fallback: use existing pooling_output logic
|
||||
if engine_core_output.pooling_output is not None:
|
||||
self._process_pooling_output(engine_core_output)
|
||||
else:
|
||||
self._process_text_output(engine_core_output)
|
||||
def process_output(self, engine_core_output):
|
||||
output_type = self._detect_output_type(engine_core_output)
|
||||
handler = self.output_handlers.get(output_type, self._process_pooling_output)
|
||||
return handler(engine_core_output)
|
||||
```
|
||||
|
||||
### 9. CLI Package Configuration
|
||||
```toml
|
||||
# pyproject.toml
|
||||
[project.scripts]
|
||||
vllm = "vllm_omni.entrypoints.cli.main:main"
|
||||
vllm-omni = "vllm_omni.entrypoints.cli.main:main"
|
||||
```
|
||||
|
||||
|
||||
## Future Work
|
||||
- Scheduler adaptations for diffusion batching.
|
||||
- Advanced output processors for audio/video.
|
||||
- Multi-worker execution (ray / multiprocess) for heavy DiT stages.
|
||||
@@ -136,4 +136,4 @@ Hongsheng Liu: [hsliuustc@gmail.com](mailto:hsliuustc@gmail.com)
|
||||
|
||||
[image10]: <data:image/png;base64,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 truncated
|
||||
|
||||
[image11]: <data:image/png;base64,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 truncated
|
||||
[image11]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABM8AAARlCAYAAACukOBGAACAAElEQVR4Xuy9B7hV13WuHSd5ch3Zia17rcQl13+K88dyXOLYlmPLRb13CVUkUEMIEB3Rkeiig+iiI3rvvffeO4gmeu+dM+56xz4TbRZ7IxDt7M33Ps98zll97rXGnGvOb4055l+YEEIIIYQQQgghhBAiJX8RXyGEEEIIIYQQQgghhEgg8UwIIYQQQgghhBBCiDRIPBNCCCGEEEIIIYQQIg0Sz4QQQgghhBBCCCGESIPEMyGEEEIIIYQQQggh0iDxTAghhBBCCCGEEEKINEg8E0IIIYQQQgghhBAiDRLPhBBCCCGEEEIIIYRIg8QzIYQQQgghhBBCCCHSIPFMCCGEEEIIIYQQQog0SDwTQgghhBBCCCGEECINEs+EuAhnz5jlKCnF09k0Kb49/B9fH98/1bazSfaXYhvr025PPm98e3xd8nJ826VcK36eVNeN/e/nih+Xajl+vaT1550jvu+lbIv/TfV/fDn5nPHz5yYhhBBCCCFE9iHxTIh05JgdP5Bjx3YrKcXSntTpeG5KXk61/lzafZFtsXPE07n1KfJz7pwp8nvB9ZL2uWBb8j7ptuWm+DWT973geqmOS943zfW+7Den+j++HD9Hym25/8eXk895wfmj8506mhOvRYQQQgghhBBZgMQzIdKBeLY/x45sz7GjO5SUvkhH0qT49vB/fH18/1TbLnbclx2bvC2+Pb4ueTm+Lb5Pqm2p9onvG///Upfj10ten25b/P/4cvzcqbalukb8nBecf7vEMyGEEEIIIbIViWdCXIQT+y3ROd6ppKSklD5RT0g8E0IIIYQQIjuReCZEOnIQz3ISHigpOstKSkpKIUk8E0IIIYQQInuReCZEOsKwTYlnSkpKX5IkngkhhBBCCJG9SDwTIh3yPFNSUrrEJPFMCCGEEEKI7EXimRDpkOeZkpLSJSaJZ0IIIYQQQmQvEs+ESIfEMyUlpUtMEs+EEEIIIYTIXiSeCZEOiWdKSkqXmCSeCSGEEEIIkb1IPBPiIpzYbxLPbmA6vjt6CEej57Dni+dgh83OHkgsu2CxL9rnpPlflk9wzIlon4Nf7HNyb2IfO56bonMe22UXXC+euD7nuZR9SX79KK/k8cuOYd/TkX3lHJKNZUOSeCaEEEIIIUT2IvFMiHRowoAbmhC8tq3bby0afWKfLd3mQhPrenToZ2MGT3ax7PRBs3lTllmJImVsxbz1dvpQ4pjypSrb8H7jfDv7LZu91t59s5gVzk1Vy9ew3RtQuOycoIZQhuCFsBYEtu1r91vfbkNs/+fsYC7K2ZGEKOYiGeuOJY4NQt6GKK9d2va0PZuOfnGu4wkB0MXAIOBFl58zabGNHDDBBTpSOD/58GXyd+ILYZB74PvkCm5nDnyxHL9/Stc3STwTQgghhBAie5F4JkQ6NGzzhiaEorWLt9hD9z9mcycvdVGJde++Ucwa1m7m4hMi1IAew+373/uBNW/Y1h/bkN6jfblRneYuPuHZNbDHCLvvrgds5viFtmDaUlsya5Xt2XjE5kxaZD079bcB3YfZwa2nXOBCdOvZaUC03zL7fNVuGzFgvK1bssUmjphhndv0sCmjZrvnG/v36NjPBvYc4fk9sSfHBa/5U5ZZwfxv2cblO23RjOXWu8tA69NlkO3//IQd3n7Gl7u172N7N+2zSmWq2aMPPm57Nh2wxTNXW6fW3WzCsOkuni2ZtSbK93Dr3qFvlI89djr6vRuWbY/y0N1mTljowtyaRZs9T3OnLPU8xe+h0vVLEs+EEEIIIYTIXiSeCXERNGzzxiWEMjzO8j39ki2YtsLFMtaVKVbeWjRs66IYAtKwvmPtyUeftVJFy9mhHcetecM2lu+pF619y09dcEPQGtpnjL30XH4XoXatP2CHo/3WR+f+nzvudE+0P995t/XpOtjGD5saXe9Fe+O1QvbgfY/YzAkLrGzxCtamWQe741f/YyXeLWP3/Pk+Wz5njbVq3N5KFStnBV9509o07eBeYORp0fSV5zzh7r/nQXvztXeiY+73/HRq3d3ee6ekVS73gQtolcpWs8ceftKFOfLfpF5Le+bxfDZ93Hx75YUC9nz025945Gmr8n4N2711X/S3ur1fspIVi84xacxUK/1eeav9QT17KV9+mzNxkZ2WgHbDksQzIYQQQgghsheJZ0JcBIlnNy5dqng2uNcoK1mkjL1fqrIN6D7CalerbzUq17HWTTu4eMZ+YwZNstv/4z8t/wsFrMDLr1vLxm1s7aLP7bmnXrAd6w5Yp1bdrH7NJlauRAUfErp1zT4r9EYRmzZmrpUrWTHa/xN7r3ApP1fZaJ9P2/exJx992lo1aW8fVKhlTz/+nB3YctLzE8SzZXPWuaj12dKt1rfrYKsV5avtx52sepS3kQNG277NB21g9+HWqM7HdmLP2SjvQ21Qr5H2wrMvW99Ph1iRQiVsanT9GePmW8WyVaPtw/28eNMd2XPUOrfuYX+68y4b2meYPfLA4+5pR/7i91Hp+iSJZ0IIIYQQQmQvEs+ESIeGbd7QhFC2bulWe+HZV1zock4kxLMOLbuee0yDeo6wmlU/cq8uvLVqVatnrZu0TwzjzPU8Q2B7u0DhRKyxXH1j3eIt7nW2c/2h6HyfutcXItzEETNt+7oDVrRQiYR4VqKifdK8k4tfDKdkqGW39n3tuadfsJpVPrKPG7T0oZOHtp2xnMNfiGcMu+TvphU7rVfHAVbnw4Z2cNux6Nge7s3W9ZNePuyzWb1WNnviYnv5+VetWf1W9uC9D/tQ0MrlPrSF01fYhOHTo99X14eSFn+39Lk4bd2jPNz9p/tcCKxfs7FNGjnTTu298D4qXZ8k8UwIIYQQQojsReKZEOmQeHZDE+LZ+mXb7Q+/v8uFq95detnMCXOtyFvF7eV8r1qvzr1sxMARLnzh/TV2yBS77Tu32bC+Y6xFo7ZWv1ZT9wTDGwvx7Df//VsX2Lq1722Deg6xuZOXWIGX37Bta/db8wZtrHHdFja490h7q0Bhq/NBA3vmiXw2bew8K/p2cR8KWrFMNc8XotqwfmPdY6xhnWY+bJLjT+xNzARKzLO3Cha2hdNX+vDPz5ZstS5te9iHlWpbuxado3O19mM/ad7Z+n06xAq88oa1b9HFr9e/21B75IHHXIwrWbSsTygwatBEK1uivK2cv87PhyhYqlhZ69Wlr4tzvToPsBLvlrbJo2b5PYvfR6XrkySeCSGEEEIIkb1IPBMiHRLPbmg6vjvH9m85YZ+262N1PmxgtarWs8G9R3nA/no1Gvsyccfw2po1YaHt3njExadd6w/ZwmkrfBbOk/uYtTPHA+03rdfSaler555pzeq3ttULNrlXF9fAW2zBtOW2d9Mxa1i7oRV45U2PZcZxU0fPskUzVvrwSfJFfLItq/b49Zi4oEGtprZpxS47tS8xacC2Nfv8vMzUOX7YNNuz8agtn7POxbrN0X51qzf0vByIrrt55S6Pp8YQT2boRFQbPWiSX2/K6Nm2dfVej82GMMaEALMmLvJhqUP6jLacI4nZOvk9vToNiOz0rN+z+H1Uuj5J4pkQQgghhBDZi8QzIdIh8eyGJ4ZJMlTTTuemI5aYZTMsn0gMy8yJ0ond5sMZmTGT4P2ITTw70sm9SceQTpqd2pc4F9dgf87F7JYli5axd94o4gIY52afMBkAeWKZ83EdO5VILIdr+frDudujv8d3J2K1cS7PB8ecTKz35ZOJ7f47o//5jZ6fQ4lznGKfQ4nze5y33PtwlGXyx/KxxO+I3z+l65ckngkhhBBCCJG9SDwTIh0Sz266hACFCIewpiGQSpeTJJ4JIYQQQgiRvUg8EyIdEs9uyoSAJi8upctNEs+EEEIIIYTIXiSeCZEOiWdKSkqXmCSeCSGEEEIIkb1IPBMiHRLPlJSULjFJPBNCCCGEECJ7kXgmRDoknikpKV1ikngmhBBCCCFE9iLxTIh0SDxTUlK6xCTxTAghhBBCiOxF4pkQ6ZB4pqSkdIlJ4pkQQgghhBDZi8QzIdIh8UxJKU8kZj89nsdnQJV4JoQQQgghRPYi8UyIdCCe7Ys6xdsTHWMlJaXrn45G6cRus5N77YJteSpF9cSpIxLPhBBCCCGEyEYkngmRhpyoH4wnCR1iJaXz0lG7cDms+7JtydtTLce3pTrvxY5LtRzfFj82+RrxY+P7xrelOzbVvun+j58nadvZE2Zd2/ewMsUr2IlDZ+x0/Nj4NeLnSZe/VMelOybVtvj/UTpzQuKZEEIIIYQQ2YjEMyGEEHmacePHWdu2bS0n52x8kxBCCCGEEEJccySeCSGEEEIIIYQQQgiRBolnQgghhBBCCCGEEEKkQeKZEEIIIYQQQgghhBBpkHgmhBAiTzNixAj76KOP7OxZxTwTQgghhBBCXH8kngkhhMjTFC9e3G699VY7ffp0fJMQQgghhBBCXHMkngkhhMjTVKxY0f7t3/7NTp06Fd8khBBCCCGEENcciWdCCCHyNJUqVbJ//dd/lXgmhBBCCCGEuCFIPBNCCJGnqVy5sv3Lv/yLxDMhhBBCCCHEDUHimRBCiDwNwzZ/+MMfSjwTQgghhBBC3BAkngkhhMjT4Hkm8UwIIYQQQghxo5B4JoQQIk+za9cuW7NmjeXk5MQ3CSGEEEIIIcQ1R+KZEEIIIYQQQgghhBBpkHgmhBBCCCGEEEIIIUQaJJ4JIYQQQgghhBBCCJEGiWdCCCHyNK1bt7Z77rlHEwYIIYQQQgghbggSz4QQQuRpqlSpYn//939vJ0+ejG8SQgghhBBCiGuOxDMhhBB5mpo1a9p3v/tdiWdCCCGEEEKIG4LEMyGEEHmaWrVq2T/+4z9KPBNCCCGEEELcECSeCSGEyNMgnv3DP/yDxDMhhBBCCCHEDUHimRBCiDxNtWrV7JZbbpF4JoQQQgghhLghSLine truncated
|
||||
@@ -12,13 +12,12 @@ vllm_omni/
|
||||
│ └── cache_config.py # Caching configurations
|
||||
├── core/ # Core processing components
|
||||
│ ├── __init__.py
|
||||
│ ├── omni_llm.py # OmniLLM and AsyncOmniLLM
|
||||
│ ├── stage_manager.py # Multi-stage orchestration
|
||||
│ ├── dit_cache_manager.py # DiT caching system
|
||||
│ └── sched/ # Schedulers
|
||||
│ ├── __init__.py
|
||||
│ ├── scheduler.py # Base scheduler interface
|
||||
│ └── diffusion_scheduler.py # DiT scheduler
|
||||
|
||||
├── engine/ # Engine components
|
||||
│ ├── __init__.py
|
||||
│ ├── processor.py # Output processing
|
||||
@@ -29,14 +28,14 @@ vllm_omni/
|
||||
│ └── diffusers_executor.py # Diffusers pipeline executor
|
||||
├── model_executor/ # Model execution components
|
||||
│ ├── __init__.py
|
||||
│ ├── ar_model_runner.py # OmniARModelRunner
|
||||
│ └── dit_model_runner.py # OmniDiffusionModelRunner
|
||||
├── worker/ # Worker implementations
|
||||
│ ├── __init__.py
|
||||
│ └── omni_worker.py # Extended worker for DiT
|
||||
│ ├── gpu_diffusion_model_runner.py # Loads and runs diffusers pipelines
|
||||
│ └── gpu_diffusion_worker.py # Worker facade around the diffusion runner
|
||||
├── entrypoints/ # Entry points and CLI
|
||||
│ ├── __init__.py
|
||||
│ ├── omni.py # OmniServeCommand
|
||||
│ ├── omni_llm.py # OmniLLM and AsyncOmniLLM
|
||||
│ └── cli/ # CLI integration
|
||||
│ ├── __init__.py
|
||||
│ └── main.py # CLI main entry point
|
||||
@@ -87,24 +86,29 @@ class OmniStageConfig:
|
||||
input_modalities: List[str]
|
||||
output_modalities: List[str]
|
||||
vllm_config: Optional[VllmConfig] = None
|
||||
executor_class: type[Executor] = MultiprocExecutor
|
||||
dit_config: Optional[DiTConfig] = None
|
||||
cache_config: Optional[DiTCacheConfig] = None
|
||||
executor_class: Optional[type[Executor]] = None
|
||||
stage_output: Optional[Any] = None
|
||||
```
|
||||
|
||||
### 3.2 DiT Configuration
|
||||
|
||||
```python
|
||||
# vllm_omni/config/dit_config.py
|
||||
# vllm_omni/config/stage_config.py
|
||||
@dataclass
|
||||
class DiTConfig:
|
||||
model_type: str
|
||||
scheduler_type: str
|
||||
model_config: Optional[ModelConfig] = None
|
||||
scheduler_config: Optional[Any] = None
|
||||
device_config: Optional[DeviceConfig] = None
|
||||
load_config: Optional[LoadConfig] = None
|
||||
compilation_config: Optional[CompilationConfig] = None
|
||||
dit_cache_config: Optional[DiTCacheConfig] = None
|
||||
num_inference_steps: int
|
||||
guidance_scale: float
|
||||
guidance_scale: float = 7.5
|
||||
use_diffusers: bool = False
|
||||
diffusers_pipeline: Optional[str] = None
|
||||
height: int = 512
|
||||
width: int = 512
|
||||
```
|
||||
|
||||
### 3.3 Cache Configuration
|
||||
@@ -179,34 +183,9 @@ class AsyncOmniLLM(AsyncLLM):
|
||||
"""Async generation interface"""
|
||||
# Similar to OmniLLM but with async/await patterns
|
||||
pass
|
||||
```
|
||||
``
|
||||
|
||||
### 4.3 DiT Scheduler Implementation
|
||||
|
||||
```python
|
||||
# vllm_omni/core/sched/diffusion_scheduler.py
|
||||
class OmniDiffusionScheduler(SchedulerInterface):
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
kv_cache_config: KVCacheConfig,
|
||||
dit_cache_config: DiTCacheConfig,
|
||||
structured_output_manager: StructuredOutputManager,
|
||||
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
|
||||
include_finished_set: bool = False,
|
||||
log_stats: bool = False,
|
||||
) -> None:
|
||||
super().__init__(vllm_config, kv_cache_config, structured_output_manager,
|
||||
mm_registry, include_finished_set, log_stats)
|
||||
self.dit_cache_manager = DiTCacheManager(dit_cache_config)
|
||||
|
||||
def schedule(self) -> SchedulerOutput:
|
||||
"""Schedule DiT requests with caching optimization"""
|
||||
# Implement DiT-specific scheduling logic
|
||||
pass
|
||||
```
|
||||
|
||||
### 4.4 Model Runners
|
||||
### 4.3 Model Runners
|
||||
|
||||
```python
|
||||
# vllm_omni/model_executor/dit_model_runner.py
|
||||
|
||||
Reference in new issue
Block a user