biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Replace FlashAttention with xformers (#70)

Browse Source
This commit is contained in:
Woosuk Kwon
2023-05-05 02:01:08 -07:00
committed by GitHub
parent 189ae23133
commit c9d5b6d4a8
13 changed files with 89 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
cacheflow/models/input_metadata.py
View File

@@ -1,6 +1,7 @@
from typing import List, Dict, Tuple
import torch
from xformers.ops.fmha.attn_bias import BlockDiagonalCausalMask
from cacheflow.sampling_params import SamplingParams
@@ -12,7 +13,6 @@ class InputMetadata:
seq_groups: List[Tuple[List[int], SamplingParams]],
seq_logprobs: Dict[int, float], # Seq id -> cumulative logprobs.
prompt_lens: List[int],
cumulative_prompt_lens: torch.Tensor,
slot_mapping: torch.Tensor,
context_lens: torch.Tensor,
max_context_len: int,
@@ -21,15 +21,14 @@ class InputMetadata:
self.seq_groups = seq_groups
self.seq_logprobs = seq_logprobs
self.prompt_lens = prompt_lens
self.cumulative_prompt_lens = cumulative_prompt_lens
self.slot_mapping = slot_mapping
self.context_lens = context_lens
self.max_context_len = max_context_len
self.block_tables = block_tables
self.attn_bias = BlockDiagonalCausalMask.from_seqlens(prompt_lens)
self.num_prompts = len(prompt_lens)
self.num_prompt_tokens = sum(prompt_lens)
self.max_prompt_len = max(prompt_lens) if prompt_lens else 0
self.num_generation_tokens = context_lens.shape[0]
self.num_valid_tokens = slot_mapping.shape[0]
if block_tables.numel() > 0:
@@ -41,15 +40,13 @@ class InputMetadata:
def __repr__(self) -> str:
return (f'InputMetadata('
f'num_prompts={self.num_prompts}, '
f'num_prompt_tokens={self.num_prompt_tokens}, '
f'max_prompt_len={self.max_prompt_len}, '
f'num_generation_tokens={self.num_generation_tokens}, '
f'num_valid_tokens={self.num_valid_tokens}, '
f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
f'max_context_len={self.max_context_len}), '
f'num_prompt_tokens={self.num_prompt_tokens}, '
f'num_prompts={self.num_prompts}, '
f'prompt_lens={self.prompt_lens}, '
f'cumulative_prompt_lens={self.cumulative_prompt_lens}, '
f'slot_mapping={self.slot_mapping}, '
f'num_generation_tokens={self.num_generation_tokens}, '
f'context_lens={self.context_lens}, '
f'block_tables={self.block_tables})')
f'max_context_len={self.max_context_len}), '
f'max_num_blocks_per_seq={self.max_num_blocks_per_seq}, '
f'block_tables={self.block_tables}), '
f'slot_mapping={self.slot_mapping}')