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.
70 lines
1.7 KiB
Python
70 lines
1.7 KiB
Python
# Adapted from
|
|
# https://github.com/modelscope/ms-swift/blob/v2.4.2/swift/utils/module_mapping.py
|
|
|
|
from dataclasses import dataclass, field
|
|
from typing import List, Union
|
|
|
|
|
|
@dataclass
|
|
class ModelKeys:
|
|
model_type: str = None
|
|
|
|
module_list: str = None
|
|
|
|
embedding: str = None
|
|
|
|
mlp: str = None
|
|
|
|
down_proj: str = None
|
|
|
|
attention: str = None
|
|
|
|
o_proj: str = None
|
|
|
|
q_proj: str = None
|
|
|
|
k_proj: str = None
|
|
|
|
v_proj: str = None
|
|
|
|
qkv_proj: str = None
|
|
|
|
qk_proj: str = None
|
|
|
|
qa_proj: str = None
|
|
|
|
qb_proj: str = None
|
|
|
|
kva_proj: str = None
|
|
|
|
kvb_proj: str = None
|
|
|
|
output: str = None
|
|
|
|
|
|
@dataclass
|
|
class MultiModelKeys(ModelKeys):
|
|
language_model: List[str] = field(default_factory=list)
|
|
connector: List[str] = field(default_factory=list)
|
|
# vision tower and audio tower
|
|
tower_model: List[str] = field(default_factory=list)
|
|
generator: List[str] = field(default_factory=list)
|
|
|
|
@staticmethod
|
|
def from_string_field(language_model: Union[str, List[str]] = None,
|
|
connector: Union[str, List[str]] = None,
|
|
tower_model: Union[str, List[str]] = None,
|
|
generator: Union[str, List[str]] = None,
|
|
**kwargs) -> 'MultiModelKeys':
|
|
|
|
def to_list(value):
|
|
if value is None:
|
|
return []
|
|
return [value] if isinstance(value, str) else list(value)
|
|
|
|
return MultiModelKeys(language_model=to_list(language_model),
|
|
connector=to_list(connector),
|
|
tower_model=to_list(tower_model),
|
|
generator=to_list(generator),
|
|
**kwargs)
|