Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Reduce block_indices and block_offsets computation to once per forward pass #102

Merged
merged 1 commit into from
Jul 17, 2024
Merged

Reduce block_indices and block_offsets computation to once per forward pass #102

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 9 additions & 3 deletions vllm/attention/backends/habana_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -142,16 +142,22 @@ def forward(
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)

if prefill_meta := attn_metadata.prefill_metadata:
block_indices = prefill_meta.block_indices
block_offsets = prefill_meta.block_offsets
if decode_meta := attn_metadata.decode_metadata:
block_indices = decode_meta.block_indices
block_offsets = decode_meta.block_offsets
if kv_cache is not None:
key_cache, value_cache = HabanaPagedAttention.split_kv_cache(
kv_cache, self.num_kv_heads, self.head_size)

# Reshape the input keys and values and store them in the cache.
# If kv_cache is not provided, the new key and value tensors are
# not cached. This happens during the initial memory profiling run.
block_indices, block_offset = cache_ops.prepare_to_cache(key_cache, attn_metadata.slot_mapping)
key_cache = self.key_cache(key, key_cache, block_indices, block_offset)
value_cache = self.value_cache(value, value_cache, block_indices, block_offset)
key_cache = self.key_cache(key, key_cache, block_indices, block_offsets)
value_cache = self.value_cache(value, value_cache, block_indices, block_offsets)

if prefill_meta := attn_metadata.prefill_metadata:
# Prompt run.
Expand Down
2 changes: 2 additions & 0 deletions vllm/attention/ops/habana_paged_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,8 @@ class HabanaPagedAttentionMetadata:
block_list: Optional[torch.Tensor]
block_mapping: Optional[torch.Tensor]
block_usage: Optional[torch.Tensor]
block_indices: Optional[torch.Tensor]
block_offsets: Optional[torch.Tensor]


class HabanaPagedAttention:
Expand Down
18 changes: 0 additions & 18 deletions vllm/hpu/cache_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,24 +10,6 @@
import habana_frameworks.torch as htorch


def reshape_and_cache(key, value, key_cache, value_cache, slot_mapping, dtype, is_prompt=False):
block_size = key_cache.size(1)
slot_mapping = slot_mapping.flatten()
indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
offsets = torch.fmod(slot_mapping, block_size)
key_cache.index_put_((indices, offsets), key)
value_cache.index_put_((indices, offsets), value)


def prepare_to_cache(cache, slot_mapping):
block_size = cache.size(1)
slot_mapping = slot_mapping.flatten()
indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
offsets = torch.fmod(slot_mapping, block_size)

return indices, offsets


def insert_or_update_cache(input, cache, block_indices, block_offsets):
cache.index_put_((block_indices, block_offsets), input)

Expand Down
19 changes: 17 additions & 2 deletions vllm/worker/habana_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,6 +198,13 @@ def pad_list(l, k, v):
return l + [v] * padding


def precompute_indices_and_offsets(block_size, slot_mapping):
slot_mapping = slot_mapping.flatten()
indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
offsets = torch.fmod(slot_mapping, block_size)
return indices, offsets


class HpuModelAdapter():
def __init__(self, model, block_size):
self.model = model
Expand Down Expand Up @@ -593,10 +600,14 @@ def _prepare_prompt(
dtype=torch.long,
device=self.device)

block_indices, block_offsets = precompute_indices_and_offsets(self.block_size, slot_mapping)

attn_metadata = self.attn_backend.make_metadata(
block_list=None,
block_mapping=None,
block_usage=None,
block_indices=block_indices,
block_offsets=block_offsets,
attn_bias=None,
seq_lens_tensor=seq_lens_tensor,
)
Expand Down Expand Up @@ -696,10 +707,14 @@ def _prepare_decode(
dtype=torch.long,
device=self.device)

block_indices, block_offsets = precompute_indices_and_offsets(self.block_size, slot_mapping)

attn_metadata = self.attn_backend.make_metadata(
block_list=block_list,
block_mapping=block_mapping,
block_usage=block_usage,
block_indices=block_indices,
block_offsets=block_offsets,
attn_bias=None,
seq_lens_tensor=None,
)
Expand Down Expand Up @@ -884,10 +899,10 @@ def _seq_len(self, attn_metadata):
def trim_attn_metadata(self, metadata: AttentionMetadata) -> object:
prefill_metadata = subtuple(metadata.prefill_metadata,
"TrimmedPrefillMetadata",
['attn_bias', 'seq_lens_tensor'])
['attn_bias', 'seq_lens_tensor', 'block_indices', 'block_offsets'])
decode_metadata = subtuple(metadata.decode_metadata,
"TrimmedDecodeMetadata",
['attn_bias', 'block_list', 'block_mapping', 'block_usage'])
['attn_bias', 'block_list', 'block_mapping', 'block_usage', 'block_indices', 'block_offsets'])
return subtuple(metadata,
'TrimmedMetadata',
['slot_mapping',
Expand Down