ef6abf3dc7
Adds ALL files needed for Dockerfile build:
- qwen3_6_scripts/ (baseline patches + our optimizations)
- vllm/ (full vllm package)
- paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
- Dockerfile + computility-run.yaml
Our optimizations vs baseline:
1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
Triton try/fallback, V2 heuristic, threshold 32K→64K
2. paged_attention_v2_pytorch.py: fills NotImplementedError,
single-bmm Phase 1 (195 launches → 3)
3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
max-num-batched-tokens 8192→16384
This repo can now be submitted to dev.modelhub.org.cn as-is.
128 lines
3.3 KiB
Python
128 lines
3.3 KiB
Python
import enum
|
|
from abc import ABC, abstractmethod
|
|
from typing import List
|
|
from typing import Sequence as GenericSequence
|
|
from typing import Tuple
|
|
|
|
from vllm.sequence import Sequence, SequenceGroup
|
|
from vllm.utils import Device
|
|
|
|
|
|
class AllocStatus(enum.Enum):
|
|
"""Result for BlockSpaceManager.can_allocate
|
|
|
|
1. Ok: seq_group can be allocated now.
|
|
2. Later: seq_group cannot be allocated.
|
|
The capacity of allocator is larger than seq_group required.
|
|
3. Never: seq_group can never be allocated.
|
|
The seq_group is too large to allocated in GPU.
|
|
"""
|
|
OK = enum.auto()
|
|
LATER = enum.auto()
|
|
NEVER = enum.auto()
|
|
|
|
|
|
class BlockSpaceManager(ABC):
|
|
|
|
@staticmethod
|
|
def get_block_space_manager_class(version: str):
|
|
version = version.lower()
|
|
|
|
if version == "v1":
|
|
from vllm.core.block_manager_v1 import BlockSpaceManagerV1
|
|
return BlockSpaceManagerV1
|
|
|
|
if version == "v2":
|
|
from vllm.core.block_manager_v2 import BlockSpaceManagerV2
|
|
return BlockSpaceManagerV2
|
|
|
|
if version == "placeholder":
|
|
from vllm.core.placeholder_block_space_manager import (
|
|
PlaceholderBlockSpaceManager)
|
|
return PlaceholderBlockSpaceManager
|
|
|
|
raise ValueError(f"Unknown version {version=}")
|
|
|
|
@abstractmethod
|
|
def can_allocate(self,
|
|
seq_group: SequenceGroup,
|
|
num_lookahead_slots: int = 0) -> AllocStatus:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def allocate(self, seq_group: SequenceGroup) -> None:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def can_append_slots(self, seq_group: SequenceGroup,
|
|
num_lookahead_slots: int) -> bool:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def append_slots(
|
|
self,
|
|
seq: Sequence,
|
|
num_lookahead_slots: int,
|
|
) -> List[Tuple[int, int]]:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def fork(self, parent_seq: Sequence, child_seq: Sequence) -> None:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def can_swap_in(self, seq_group: SequenceGroup,
|
|
num_lookahead_slots: int) -> AllocStatus:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def swap_in(self, seq_group: SequenceGroup) -> List[Tuple[int, int]]:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def can_swap_out(self, seq_group: SequenceGroup) -> bool:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def swap_out(self, seq_group: SequenceGroup) -> List[Tuple[int, int]]:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def free(self, seq: Sequence) -> None:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def get_block_table(self, seq: Sequence) -> List[int]:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def get_num_free_gpu_blocks(self) -> int:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def get_num_free_cpu_blocks(self) -> int:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def access_all_blocks_in_seq(
|
|
self,
|
|
seq: Sequence,
|
|
access_time: float,
|
|
) -> None:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def get_common_computed_block_ids(
|
|
self, seqs: List[Sequence]) -> GenericSequence[int]:
|
|
pass
|
|
|
|
@abstractmethod
|
|
def mark_blocks_as_computed(self, seq_group: SequenceGroup,
|
|
token_chunk_size: int):
|
|
pass
|
|
|
|
@abstractmethod
|
|
def get_prefix_cache_hit_rate(self, device: Device) -> float:
|
|
"""Prefix cache hit rate. -1 means not supported or disabled."""
|
|
pass
|