44 lines
1.7 KiB
Python
44 lines
1.7 KiB
Python
import os
|
|
from dataclasses import dataclass
|
|
from transformers import AutoConfig
|
|
|
|
|
|
@dataclass
|
|
class Config:
|
|
model: str
|
|
max_num_batched_tokens: int = 16384
|
|
max_num_seqs: int = 512
|
|
max_model_len: int = 4096
|
|
gpu_memory_utilization: float = 0.9
|
|
tensor_parallel_size: int = 1
|
|
enforce_eager: bool = False
|
|
hf_config: AutoConfig | None = None
|
|
eos: int = -1
|
|
kvcache_block_size: int = 4096
|
|
num_kvcache_blocks: int = -1
|
|
|
|
# CPU Offload configuration
|
|
enable_cpu_offload: bool = False
|
|
offload_policy: str = "lru" # "lru", "fifo", or full class path
|
|
num_transfer_streams: int = 4 # Number of CUDA streams for async transfers
|
|
num_gpu_blocks: int = -1 # User-specified GPU blocks count, -1 = auto (use max available)
|
|
|
|
# Computed fields for offload (set in __post_init__ or by ModelRunner)
|
|
num_gpu_kvcache_blocks: int = -1
|
|
num_cpu_kvcache_blocks: int = -1
|
|
|
|
# Sparse attention configuration
|
|
sparse_policy: str | None = None # "vertical_slash", "quest", "streaming_llm", or None
|
|
sparse_num_sink_blocks: int = 1 # Number of sink blocks for sparse patterns
|
|
sparse_local_window_blocks: int = 2 # Local window size for VerticalSlash
|
|
sparse_topk_blocks: int = 8 # Top-K blocks for Quest
|
|
sparse_threshold_blocks: int = 4 # Apply sparse only when blocks > threshold
|
|
|
|
def __post_init__(self):
|
|
assert os.path.isdir(self.model)
|
|
assert self.kvcache_block_size % 256 == 0
|
|
assert 1 <= self.tensor_parallel_size <= 8
|
|
self.hf_config = AutoConfig.from_pretrained(self.model)
|
|
self.max_model_len = min(self.max_model_len, self.hf_config.max_position_embeddings)
|
|
assert self.max_num_batched_tokens >= self.max_model_len
|