These 17 commits are when the Protocol Buffers files have changed:
| Commit: | 9242ebd | |
|---|---|---|
| Author: | Koa Calloway | |
Consolidate framework-level node types into DATA_OP_NODE. Use op_info oneof to store either StorageInfo or DataOpType.
The documentation is generated from this commit.
| Commit: | 4326f41 | |
|---|---|---|
| Author: | Dan Mihailescu | |
Drop parent/child node-type comments from NodeType enum Remove the comments coupling STORAGE_NODE to the framework-level parent nodes: tracing may capture only framework-level nodes, or only storage operations, so the asserted relationship is not mandatory. Per review feedback. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| Commit: | ccf0ef5 | |
|---|---|---|
| Author: | Dan Mihailescu | |
Move storage operations into storage.proto Extract StorageLevel, StorageInfo, FsOp, PageOp, and BlockOp into a new storage.proto and import it from et_def.proto, per review feedback (#171). protoc's --python_out emits flat absolute imports between generated modules (et_def_pb2 contains `import storage_pb2`), which fail under the installed chakra.schema.protobuf package. Add schema/protobuf/__init__.py that puts the package directory on sys.path so the generated cross-imports resolve. See protocolbuffers/protobuf#1491. Register storage.proto in setup.cfg (proto_files + grpc_files) and pyproject package-data so it is compiled and shipped alongside et_def.proto. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| Commit: | b21f4ea | |
|---|---|---|
| Author: | Dan Mihailescu | |
Rename inner Op enum to OpType in storage op messages Rename the nested Op enum to OpType in FsOp, PageOp, and BlockOp so it no longer reads ambiguously next to the FsOp message name, per review feedback. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| Commit: | 3126df4 | |
|---|---|---|
| Author: | Dan Mihailescu | |
| Committer: | Dan Mihailescu | |
Rework storage node types into framework-level parents and storage ops Rename BATCH_NODE to DATALOADER_NODE and add KV_TRANSFER_NODE so dataloader, checkpoint, and KV-transfer ops each have a distinct framework-level parent node type, per review feedback. STORAGE_NODE remains the lower-level storage operation emitted under those parents. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| Commit: | 86326cd | |
|---|---|---|
| Author: | Koa Calloway | |
| Committer: | Dan Mihailescu | |
Add storage operations to et_def.proto
| Commit: | a1120dd | |
|---|---|---|
| Author: | Taekyung Heo | |
| Committer: | Taekyung Heo | |
Reorganize directories
| Commit: | a2b45c1 | |
|---|---|---|
| Author: | Taekyung Heo | |
[chakra][et_def] Add version field Summary: Add version field Reviewed By: shengbao Differential Revision: D50082609
| Commit: | c236cfa | |
|---|---|---|
| Author: | Taekyung Heo | |
[chakra][et_def] Update the Chakra schema Summary: * Transition from proto2 to proto3 * Removed "required" or "optional" keywords * Updated AttributeProto to incorporate all data types in proto3 * Renamed parent to data_dep & ctrl_dep * Updated CollectiveCommType so that it can incorporate all collective communication types * Added "Message Tensor" to encode tensor information * Updated the input / output data types to message IOInfo * Added operator start time and duration Reviewed By: lofe Differential Revision: D48790087
| Commit: | 29de3e1 | |
|---|---|---|
| Author: | Taekyung Heo | |
[chakra] support broadcast collective communication Summary: Support broadcast collective communication Reviewed By: shengbao Differential Revision: D48032038
| Commit: | 7bab792 | |
|---|---|---|
| Author: | Taekyung Heo | |
[chakra][et_def] addition of uint64 to AttributeProto Summary: In this commit, I have extended the data types of AttributeProto to include uint64. Prior to this change, AttributeProto lacked support for the uint64 data type. Reviewed By: shengbao Differential Revision: D47952700
| Commit: | b03abd2 | |
|---|---|---|
| Author: | Taekyung Heo | |
[chakra][et_def] add GlobalMetadata to track global metadata Summary: While PyTorch execution traces have global metadata such as schema, pid, time, and start_ts, Chakra execution traces do not. I have introduced the ExecutionTrace message to track global metadata. Reviewed By: zhaodongwang Differential Revision: D47795237
| Commit: | 7f6b398 | |
|---|---|---|
| Author: | Taekyung Heo | |
| Committer: | Taekyung Heo | |
[chakra][et_def] change 'strings' field type from 'bytes' to 'string' in AttributeProto Summary: Updated the data type in AttributeProto: Changed 'strings' field from 'repeated bytes' to 'repeated string'. Reviewed By: lofe Differential Revision: D47636929
| Commit: | 9d917e3 | |
|---|---|---|
| Author: | Taekyung Heo | |
[et_def] update schema to v0.0.3
| Commit: | 2719463 | |
|---|---|---|
| Author: | Taekyung Heo | |
| Committer: | Taekyung Heo | |
[et_def] update schema to v0.0.2
| Commit: | ec3832c | |
|---|---|---|
| Author: | Taekyung Heo | |
[et_def] rename eg_def to et_def
| Commit: | df405e8 | |
|---|---|---|
| Author: | Taekyung Heo | |
| Committer: | Taekyung Heo | |
[eg_def] add execution graph definition