Skip to content

bugfix: ensure tensor contiguous layout before protobuf serialization. #1468

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
a120092009 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

bugfix: ensure tensor contiguous layout before protobuf serialization. #1468

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
42b7646
bugfix: ensure tensor contiguous layout before protobuf serialization.
a120092009 May 18, 2026
bc833c8
refactor: move contiguous call into bfloat16 and half serialization b…
a120092009 May 18, 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
18 changes: 12 additions & 6 deletions xllm/core/util/utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -567,25 +567,31 @@ bool torch_to_proto(const torch::Tensor& torch_tensor,
proto_contents, torch_tensor, proto_datatype);
break;
case torch::kBFloat16: {
// Need to convert bfloat16 to uint8_t for storage
auto bfloat16_ptr = torch_tensor.data_ptr<torch::BFloat16>();
// Need to convert bfloat16 to uint8_t for storage.
// Ensure contiguous layout first — data_ptr() reads raw memory
// which would be wrong for NHWC/ChannelsLast tensors.
torch::Tensor contig_tensor = torch_tensor.contiguous();
auto bfloat16_ptr = contig_tensor.data_ptr<torch::BFloat16>();
Comment thread
a120092009 marked this conversation as resolved.
uint8_t* uint8_ptr = reinterpret_cast<uint8_t*>(bfloat16_ptr);
torch::Tensor uint8_tensor =
torch::from_blob(uint8_ptr,
{static_cast<int64_t>(torch_tensor.numel() *
{static_cast<int64_t>(contig_tensor.numel() *
sizeof(torch::BFloat16))},
torch::dtype(torch::kUInt8));
data_set_success = set_data_to_contents<uint8_t>(
proto_contents, uint8_tensor, proto_datatype);
break;
}
case torch::kHalf: {
// Need to convert float16 to uint8_t for storage
auto float16_ptr = torch_tensor.data_ptr<torch::Half>();
// Need to convert float16 to uint8_t for storage.
// Ensure contiguous layout first — data_ptr() reads raw memory
// which would be wrong for NHWC/ChannelsLast tensors.
torch::Tensor contig_tensor = torch_tensor.contiguous();
auto float16_ptr = contig_tensor.data_ptr<torch::Half>();
Comment thread
a120092009 marked this conversation as resolved.
uint8_t* uint8_ptr = reinterpret_cast<uint8_t*>(float16_ptr);
torch::Tensor uint8_tensor = torch::from_blob(
uint8_ptr,
{static_cast<int64_t>(torch_tensor.numel() * sizeof(torch::Half))},
{static_cast<int64_t>(contig_tensor.numel() * sizeof(torch::Half))},
torch::dtype(torch::kUInt8));
data_set_success = set_data_to_contents<uint8_t>(
proto_contents, uint8_tensor, proto_datatype);
Expand Down
Loading