[Mistral] Mistral-7B-v0.1 support (#1196)
Co-authored-by: timlacroix <t@mistral.ai>
This commit is contained in:
co-authored by
timlacroix
parent
7bedab5748
commit
bb1ba58f06
@@ -1,4 +1,4 @@
|
||||
from typing import Dict, List, Tuple
|
||||
from typing import Dict, List, Optional, Tuple
|
||||
|
||||
import torch
|
||||
from xformers.ops import AttentionBias
|
||||
@@ -29,6 +29,7 @@ class InputMetadata:
|
||||
context_lens: torch.Tensor,
|
||||
max_context_len: int,
|
||||
block_tables: torch.Tensor,
|
||||
sliding_window: Optional[int] = None,
|
||||
) -> None:
|
||||
self.seq_groups = seq_groups
|
||||
self.seq_data = seq_data
|
||||
@@ -38,6 +39,24 @@ class InputMetadata:
|
||||
self.max_context_len = max_context_len
|
||||
self.block_tables = block_tables
|
||||
|
||||
self.to_cache = None
|
||||
if sliding_window is not None:
|
||||
# We need to keep the positions of sliding windows within
|
||||
# the key / value tables, this is helpful to know which
|
||||
# elements we need to cache and where
|
||||
to_cache, start_idx = [], 0
|
||||
for prompt_len in self.prompt_lens:
|
||||
to_cache.extend(
|
||||
range(
|
||||
start_idx + max(0, prompt_len - sliding_window),
|
||||
start_idx + prompt_len,
|
||||
))
|
||||
start_idx += prompt_len
|
||||
to_cache.extend(range(start_idx, slot_mapping.shape[0]))
|
||||
self.to_cache = torch.tensor(to_cache,
|
||||
dtype=torch.int32,
|
||||
device=self.slot_mapping.device)
|
||||
|
||||
self.num_prompts = len(prompt_lens)
|
||||
self.num_prompt_tokens = sum(prompt_lens)
|
||||
self.num_generation_tokens = context_lens.shape[0]
|
||||
|
||||
Reference in New Issue
Block a user