Skip to content

[PD] reduce a D2D copy. #1462

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
wants to merge 2 commits into
base: deepseek_r1
Choose a base branch
from
Open

[PD] reduce a D2D copy. #1462

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
123e47e
try remove h2d, direct write cpu tensor to kv cache
jikunshang Jun 20, 2025
07cebb1
fix
jikunshang Jun 23, 2025
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
11 changes: 6 additions & 5 deletions vllm/distributed/kv_transfer/kv_connector/mooncake_store_connector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,9 +330,9 @@ def recv_kv_caches_and_hidden_states_hpu(
block_indices_tensor = torch.tensor(block_indices_list[start_block_idx:end_block_idx], device="hpu", dtype=torch.int32 )
# we think this is a padding sequence, so we skip it. but we still need write kv cache
if slen == 1:
for i in range(model_executable.model.model.start_layer,
model_executable.model.model.end_layer):
current_layer_idx = i - model_executable.model.model.start_layer
for i in range(model_executable.model.start_layer,
model_executable.model.end_layer):
current_layer_idx = i - model_executable.model.start_layer
kv_cache = kv_caches[current_layer_idx]
# key_cache, value_cache = kv_cache[0], kv_cache[1]
key_cache = kv_cache[0]
Expand Down Expand Up @@ -371,8 +371,9 @@ def recv_kv_caches_and_hidden_states_hpu(
num_computed_tokens_list.append(num_computed_tokens)

# it's padded to block size now.
key_values = remote_kv.to("hpu")
keys = key_values
# key_values = remote_kv.to("hpu")
# TEST: use CPU kv cache directly
keys = remote_kv
# values = key_values[..., self.k_head_size:]

htorch.core.mark_step()
Expand Down
Loading