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.
40 lines
1.2 KiB
Python
40 lines
1.2 KiB
Python
"""Utilities for selecting and loading models."""
|
|
import contextlib
|
|
from typing import Tuple, Type
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from vllm.config import ModelConfig
|
|
from vllm.model_executor.models import ModelRegistry
|
|
|
|
|
|
@contextlib.contextmanager
|
|
def set_default_torch_dtype(dtype: torch.dtype):
|
|
"""Sets the default torch dtype to the given dtype."""
|
|
old_dtype = torch.get_default_dtype()
|
|
torch.set_default_dtype(dtype)
|
|
yield
|
|
torch.set_default_dtype(old_dtype)
|
|
|
|
|
|
def get_model_architecture(
|
|
model_config: ModelConfig) -> Tuple[Type[nn.Module], str]:
|
|
architectures = getattr(model_config.hf_config, "architectures", [])
|
|
# Special handling for quantized Mixtral.
|
|
# FIXME(woosuk): This is a temporary hack.
|
|
mixtral_supported = [
|
|
"fp8", "compressed-tensors", "gptq_marlin", "awq_marlin"
|
|
]
|
|
|
|
if (model_config.quantization is not None
|
|
and model_config.quantization not in mixtral_supported
|
|
and "MixtralForCausalLM" in architectures):
|
|
architectures = ["QuantMixtralForCausalLM"]
|
|
|
|
return ModelRegistry.resolve_model_cls(architectures)
|
|
|
|
|
|
def get_architecture_class_name(model_config: ModelConfig) -> str:
|
|
return get_model_architecture(model_config)[1]
|