Skip to content

A5 support reshape and cache in CP situation #7636

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

Open
lenghuixing0330 wants to merge 8 commits into
base: main
Choose a base branch
from
+10 −5
Open

A5 support reshape and cache in CP situation #7636

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
0ba8c4b
A5 reshape and cache
lenghuixing0330 Mar 25, 2026
0bcf2de
clean code
lenghuixing0330 Mar 25, 2026
89b43d7
clean code 2
lenghuixing0330 Mar 25, 2026
7aee4fa
clean code 3
lenghuixing0330 Mar 25, 2026
7932d80
Merge branch 'main' into scatter_fix
lenghuixing0330 Mar 26, 2026
27abc2e
clean code 4
lenghuixing0330 Mar 26, 2026
f0a0bd5
Merge branch 'scatter_fix' of https://github.com/lenghuixing0330/vllm…
lenghuixing0330 Mar 26, 2026
734b576
clean code 5
lenghuixing0330 Mar 26, 2026
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
9 changes: 5 additions & 4 deletions vllm_ascend/attention/context_parallel/attention_cp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,6 +51,7 @@
split_decodes_and_prefills,
)
from vllm_ascend.compilation.acl_graph import get_graph_params, update_graph_params_workspaces
from vllm_ascend.device.device_op import DeviceOperator
from vllm_ascend.utils import cp_chunkedprefill_comm_stream, weak_ref_tensors


Expand Down Expand Up @@ -759,12 +760,12 @@ def reshape_and_cache(

if has_decode:
slot_mapping = attn_metadata.slot_mapping[: num_decode_tokens * self.pcp_size : self.pcp_size]
torch_npu._npu_reshape_and_cache(
DeviceOperator.reshape_and_cache(
key=key[:num_decode_tokens],
value=value[:num_decode_tokens],
key_cache=self.key_cache,
value_cache=self.value_cache,
slot_indices=slot_mapping,
slot_mapping=slot_mapping,
)

if has_prefill:
Expand All @@ -791,12 +792,12 @@ def reshape_and_cache(
slot_mapping = attn_metadata.slot_mapping[
self.pcp_size * num_decode_tokens : attn_metadata.num_actual_tokens_pcp_padded
]
torch_npu._npu_reshape_and_cache(
DeviceOperator.reshape_and_cache(
key=prefill_key,
value=prefill_value,
key_cache=self.key_cache,
value_cache=self.value_cache,
slot_indices=slot_mapping,
slot_mapping=slot_mapping,
)
if self.is_kv_producer:
attn_metadata.reshape_cache_event.record()
Expand Down
6 changes: 5 additions & 1 deletion vllm_ascend/device/device_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,7 +204,11 @@ class A5DeviceAdaptor(BaseDeviceAdaptor):
@classmethod
def reshape_and_cache(cls, key, value, key_cache, value_cache, slot_mapping):
torch_npu.npu_scatter_pa_kv_cache(
key=key, value=value.contiguous(), key_cache=key_cache, value_cache=value_cache, slot_mapping=slot_mapping
key=key.contiguous(),
value=value.contiguous(),
key_cache=key_cache,
value_cache=value_cache,
slot_mapping=slot_mapping.contiguous(),
)

@staticmethod
Expand Down
Loading