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.
50 lines
1.5 KiB
Python
50 lines
1.5 KiB
Python
import os
|
|
from typing import Optional, Union
|
|
|
|
from transformers import AutoConfig, PretrainedConfig
|
|
|
|
|
|
class EAGLEConfig(PretrainedConfig):
|
|
model_type = "eagle"
|
|
|
|
def __init__(self,
|
|
model: Union[PretrainedConfig, dict, None] = None,
|
|
truncated_vocab_size: Optional[int] = None,
|
|
**kwargs):
|
|
|
|
model_config = None if model is None else (AutoConfig.for_model(
|
|
**model) if isinstance(model, dict) else model)
|
|
|
|
for k, v in kwargs.items():
|
|
if k != "architectures" and k != "model_type" and hasattr(
|
|
model_config, k):
|
|
setattr(model_config, k, v)
|
|
|
|
self.model = model_config
|
|
|
|
if self.model is None:
|
|
self.truncated_vocab_size = None
|
|
else:
|
|
self.truncated_vocab_size = self.model.vocab_size if \
|
|
truncated_vocab_size is None else truncated_vocab_size
|
|
|
|
if "architectures" not in kwargs:
|
|
kwargs["architectures"] = ["EAGLEModel"]
|
|
|
|
super().__init__(**kwargs)
|
|
|
|
if self.model is not None:
|
|
for k, v in self.model.to_dict().items():
|
|
if not hasattr(self, k):
|
|
setattr(self, k, v)
|
|
|
|
@classmethod
|
|
def from_pretrained(
|
|
cls,
|
|
pretrained_model_name_or_path: Union[str, os.PathLike],
|
|
**kwargs,
|
|
) -> "EAGLEConfig":
|
|
config_dict, kwargs = cls.get_config_dict(
|
|
pretrained_model_name_or_path, **kwargs)
|
|
return cls.from_dict(config_dict, **kwargs)
|