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.
18 lines
583 B
Python
18 lines
583 B
Python
from vllm.envs import VLLM_USE_MODELSCOPE
|
|
|
|
if VLLM_USE_MODELSCOPE:
|
|
# Patch here, before each import happens
|
|
import modelscope
|
|
from packaging import version
|
|
|
|
# patch_hub begins from modelscope>=1.18.1
|
|
if version.parse(modelscope.__version__) <= version.parse('1.18.0'):
|
|
raise ImportError(
|
|
'Using vLLM with ModelScope needs modelscope>=1.18.1, please '
|
|
'install by `pip install modelscope>=1.18.1`')
|
|
|
|
from modelscope.utils.hf_util import patch_hub
|
|
|
|
# Patch hub to download models from modelscope to speed up.
|
|
patch_hub()
|