package itex

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: GraphOptions

• string allowlist_add = 1

  AutoMixedPrecisionOptions. For AutoMixedPrecision, we has four lists: allowlist, denylist, inferlist and clearlist, add_* add ops corresponding list. but you must remove these ops from original list. for example. AvgPool in InferList. if you add AvgPool to AllowList, you should add_allowlist = "AvgPool" and remove_inferlist = "AvgPool".

• string inferlist_add = 2

• string clearlist_add = 3

• string denylist_add = 4

• string allowlist_remove = 5

• string inferlist_remove = 6

• string clearlist_remove = 7

• string denylist_remove = 8

• bool unsafe_force_all = 9

  convert all float32 ops to float16/bfloat16. now, only support float16 data type.

• ITEXDataType data_type = 10

  Set data type for AutoMixedPrecision.

• optional GraphOptions graph_options = 1

• optional DebugOptions debug_options = 2

Used in: ConfigProto

• string auto_mixed_precision_log_path = 1

  Save auto mixed precision "pre-optimization" and "post-optimization" graph to log path.

• Toggle xpu_force_sync = 2

  Run the graph with sync mode (default is OFF).

Used in: ConfigProto

• Toggle onednn_graph = 1

  (default is OFF).

• Toggle layout_opt = 2

  Optimize onednn layouts (default is ON). A graph pass that rewrites graph for propagating oneDNN layout as a tensor.

• Toggle remapper = 3

  Remapping (default is ON) Remap subgraphs onto more efficient implementations.

• Toggle auto_mixed_precision = 4

  Optimize data types for ITEX (default is OFF). This will try to use bfloat16 on CPUs and use Bfloat16/Float16 on GPUs, which is faster. Note that this can change the numerical stability of the graph.

• optional AutoMixedPrecisionOptions auto_mixed_precision_options = 5

• Toggle sharding = 6

  Auto sharding (default is OFF) Shard single device graph on multiple devices.

• optional ShardingConfig sharding_config = 7

• bool device_isxehpc = 8

  Get device isXeHPC()

• bool device_hasxmx = 9

  Get device HasXMX()

Used in: AutoMixedPrecisionOptions

• DEFAULT_DATA_TYPE = 0

• FLOAT16 = 1

• BFLOAT16 = 2

Used in: GraphOptions

• bool auto_mode = 1

• repeated ShardingDevices devices = 2

  A list of devices and their configs to run auto sharding.

Used in: ShardingConfig

• string device_type = 1

  Automatically select best config based on cost model.

• int32 device_num = 2

  Device number for each device type that is allowed for auto sharding.

• int32 batch_size = 3

  Batch size on each device.

• int32 stage_num = 4

  Accumulate stage number is used for gradient accumulation, which means running a configured number of steps without updating the model variables while accumulating the gradients of those steps and then using the accumulated gradients to compute the variable updates.

Used in: DebugOptions, GraphOptions

• DEFAULT = 0

• ON = 1

• OFF = 2