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

Jump to bottom

adapt npu #11716

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open

adapt npu #11716

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
890765a
test
yangw1234 Jul 30, 2024
259c01a
adapt npu
yangw1234 Aug 6, 2024
ce28bcc
use int4 to run lm_head
yangw1234 Aug 7, 2024
3e2d17a
clean up
yangw1234 Aug 9, 2024
429cf53
support transpose value
yangw1234 Aug 13, 2024
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
Prev Previous commit
support transpose value
  • Loading branch information
@yangw1234
yangw1234 committed Aug 13, 2024
commit 429cf53e506d136dcad27f4156449f9a1a85b849
9 changes: 5 additions & 4 deletions python/llm/src/ipex_llm/transformers/kv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,8 +82,8 @@ def update(

batch_size, num_heads, seq_len, head_dim = key_states.shape

max_seq_length = cache_kwargs.pop("max_seq_len", None)
transpose_value = cache_kwargs.pop("transpose_value", None)
max_seq_length = cache_kwargs["max_seq_len"] if "max_seq_len" in cache_kwargs else None
transpose_value = cache_kwargs["transpose"] if "transpose" in cache_kwargs else False

if layer_idx == 0 or layer_idx == 16:
if hasattr(self, "_seen_tokens"):
Expand All @@ -101,8 +101,9 @@ def update(
batch_size, num_heads, head_dim,
0, max_len,
key_states.dtype, key_states.device,
tranpose_value=transpose_value
)
k_cache, v_cache = append_kv_cache(k_cache, v_cache, key_states, value_states)
k_cache, v_cache = append_kv_cache(k_cache, v_cache, key_states, value_states, transpose_value=transpose_value)

self.key_cache[layer_idx] = k_cache
self.value_cache[layer_idx] = v_cache
Expand All @@ -122,7 +123,7 @@ def update(
new_v_cache[...] = v_cache[...]
k_cache = new_k_cache
v_cache = new_v_cache
k_cache, v_cache = append_kv_cache(k_cache, v_cache, key_states, value_states)
k_cache, v_cache = append_kv_cache(k_cache, v_cache, key_states, value_states, transpose_value=transpose_value)
self.key_cache[layer_idx] = k_cache
self.value_cache[layer_idx] = v_cache

Expand Down
87 changes: 62 additions & 25 deletions python/llm/src/ipex_llm/transformers/models/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,23 +37,41 @@ def decoding_fast_path_qtype_check(proj):
return qtype in [SYM_INT4, FP8E5, FP4]


def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
key_cache_storage = torch.zeros(batch_size, num_heads,
max_length, head_dim,
dtype=dtype, device=device)
value_cache_storage = torch.zeros(batch_size, num_heads,
max_length, head_dim,
dtype=dtype, device=device)

key_cache = key_cache_storage.as_strided((batch_size, num_heads,
current_length, head_dim),
key_cache_storage.stride(),
storage_offset=0)
value_cache = value_cache_storage.as_strided((batch_size, num_heads,
def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device, tranpose_value=False):
if not tranpose_value:
key_cache_storage = torch.zeros(batch_size, num_heads,
max_length, head_dim,
dtype=dtype, device=device)
value_cache_storage = torch.zeros(batch_size, num_heads,
max_length, head_dim,
dtype=dtype, device=device)

key_cache = key_cache_storage.as_strided((batch_size, num_heads,
current_length, head_dim),
value_cache_storage.stride(),
key_cache_storage.stride(),
storage_offset=0)
return key_cache, value_cache
value_cache = value_cache_storage.as_strided((batch_size, num_heads,
current_length, head_dim),
value_cache_storage.stride(),
storage_offset=0)
return key_cache, value_cache
else:
key_cache_storage = torch.zeros(batch_size, num_heads,
max_length, head_dim,
dtype=dtype, device=device)
value_cache_storage = torch.zeros(batch_size, num_heads,
head_dim, max_length,
dtype=dtype, device=device)

key_cache = key_cache_storage.as_strided((batch_size, num_heads,
current_length, head_dim),
key_cache_storage.stride(),
storage_offset=0)
value_cache = value_cache_storage.as_strided((batch_size, num_heads,
head_dim, current_length),
value_cache_storage.stride(),
storage_offset=0)
return key_cache, value_cache.transpose(-1, -2)


def extend_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
Expand All @@ -63,16 +81,35 @@ def extend_kv_cache(batch_size, num_heads, head_dim, current_length, max_length,
return init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device)


def append_kv_cache(cache_k, cache_v, key_states, value_states):
new_size = (cache_k.size(0),
cache_k.size(1),
cache_k.size(2) + key_states.size(2),
cache_k.size(3))
new_cache_k = cache_k.as_strided(new_size, cache_k.stride(), storage_offset=0)
new_cache_k[:, :, cache_k.size(2):cache_k.size(2) + key_states.size(2), :] = key_states
new_cache_v = cache_v.as_strided(new_size, cache_v.stride(), storage_offset=0)
new_cache_v[:, :, cache_v.size(2):cache_v.size(2) + key_states.size(2), :] = value_states
return new_cache_k, new_cache_v
def append_kv_cache(cache_k, cache_v, key_states, value_states, transpose_value=False):
if not transpose_value:
new_size = (cache_k.size(0),
cache_k.size(1),
cache_k.size(2) + key_states.size(2),
cache_k.size(3))
new_cache_k = cache_k.as_strided(new_size, cache_k.stride(), storage_offset=0)
new_cache_k[:, :, cache_k.size(2):cache_k.size(2) + key_states.size(2), :] = key_states
new_cache_v = cache_v.as_strided(new_size, cache_v.stride(), storage_offset=0)
new_cache_v[:, :, cache_v.size(2):cache_v.size(2) + key_states.size(2), :] = value_states
return new_cache_k, new_cache_v
else:
new_size_key = (cache_k.size(0),
cache_k.size(1),
cache_k.size(2) + key_states.size(2),
cache_k.size(3))
new_cache_k = cache_k.as_strided(new_size_key, cache_k.stride(), storage_offset=0)
new_cache_k[:, :, cache_k.size(2):cache_k.size(2) + key_states.size(2), :] = key_states

new_size_value = (cache_v.size(0),
cache_v.size(1),
cache_v.size(3),
cache_v.size(2) + value_states.size(3),
)
raw_cache_v = cache_v.transpose(-1, -2)
# assert raw_cache_v.is_contiguous(), f"raw_cache_v size is {raw_cache_v.shape}, stride is {raw_cache_v.stride()}"
new_cache_v = raw_cache_v.as_strided(new_size_value, raw_cache_v.stride(), storage_offset=0)
new_cache_v[:, :, :, raw_cache_v.size(3):raw_cache_v.size(3) + value_states.size(3)] = value_states
return new_cache_k, new_cache_v.transpose(-1, -2)


def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor) -> bool:
Expand Down