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
2.0 KiB
Python
70 lines
2.0 KiB
Python
from dataclasses import dataclass
|
|
from typing import Any, Dict, List, Tuple
|
|
|
|
import torch
|
|
|
|
from vllm.pooling_params import PoolingParams
|
|
from vllm.utils import is_pin_memory_available
|
|
|
|
|
|
class PoolingMetadata:
|
|
"""Metadata for pooling operations in the Pooler layer.
|
|
|
|
This class holds the necessary information for pooling operations,
|
|
providing context for how to perform pooling and other related operations.
|
|
|
|
Attributes:
|
|
seq_groups: List of (seq_ids, pooling_params).
|
|
seq_data: A mapping of sequence ID to additional sequence data.
|
|
prompt_lens: List of the lengths of each prompt.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
seq_groups: List[Tuple[List[int], PoolingParams]],
|
|
seq_data: Dict[int, Any], # Specific data related to sequences
|
|
prompt_lens: List[int],
|
|
) -> None:
|
|
self.seq_groups = seq_groups
|
|
self.seq_data = seq_data
|
|
self.prompt_lens = prompt_lens
|
|
|
|
def __repr__(self) -> str:
|
|
return ("PoolingMetadata("
|
|
f"seq_groups={self.seq_groups}, "
|
|
f"seq_data={self.seq_data}, "
|
|
f"prompt_lens={self.prompt_lens})")
|
|
|
|
|
|
@dataclass
|
|
class PoolingTensors:
|
|
"""Tensors for pooling."""
|
|
|
|
prompt_lens: torch.Tensor
|
|
|
|
@classmethod
|
|
def from_pooling_metadata(
|
|
cls,
|
|
pooling_metadata: "PoolingMetadata",
|
|
device: torch.device,
|
|
) -> "PoolingTensors":
|
|
"""
|
|
Create PoolingTensors from PoolingMetadata.
|
|
|
|
Args:
|
|
pooling_metadata: PoolingMetadata instance to convert.
|
|
device: Device to store the tensors.
|
|
"""
|
|
# Convert prompt lengths to tensor
|
|
pin_memory = is_pin_memory_available()
|
|
|
|
prompt_lens_t = torch.tensor(
|
|
pooling_metadata.prompt_lens,
|
|
device="cpu",
|
|
dtype=torch.long,
|
|
pin_memory=pin_memory,
|
|
)
|
|
|
|
return cls(prompt_lens=prompt_lens_t.to(device=device,
|
|
non_blocking=True), )
|