Add on-the-fly bfloat16->float16 conversion pass #740
Conversation
| @@ -110,7 +111,7 @@ struct ProviderInfo { | |||
| const ConfigOptions* config_options{NULL}; | |||
| const std::unordered_set<std::string> valid_provider_keys = {"device_type", "device_id", "device_luid", "cache_dir", "precision", | |||
| "load_config", "context", "num_of_threads", "model_priority", "num_streams", "enable_opencl_throttling", "enable_qdq_optimizer", | |||
| "enable_causallm", "disable_dynamic_shapes", "reshape_input"}; | |||
| "enable_bfloat16_optimizer", "enable_causallm", "disable_dynamic_shapes", "reshape_input"}; | |||
There was a problem hiding this comment.
Why is a separate provider option needed, is it possible to detect the model has bfloat16 datatype and intrinsically enable optimization ?
There was a problem hiding this comment.
@sfatimar, because at some point OpenVINO might enable the native execution of bfloat16 models. This is a workaround until this functionality is enabled. Let's discuss it with Mayuresh and act accordingly.
Regarding the graph optimizations link you've shared: strictly speaking, it's not the same kind of optimizations as are those in the list. It's more like the QDQ scales fix we implemented earlier.
There was a problem hiding this comment.
We cannot have external provider options as workaround because it impacts external users and apps and need to be given a deprecation notice 2 releases in advance. I would prefer it to be handled internally
There was a problem hiding this comment.
Please check with Mayuresh if ProviderOption can be avoided by intrinsically detecting BFloat16 or adding ep specific optimization pass https://onnxruntime.ai/docs/performance/model-optimizations/graph-optimizations.html#extended-graph-optimizations
|
Tests are prepared and pushed to a separate branch until we have a confirmation whether we want them in onnxruntime or in the internal testing repo |
This PR is adding a functionality to convert bfloat16 models to float16 models. It's using a lot of functionality from the QDQ scales fix introduced recently.
To be added:
https://jira.devtools.intel.com/browse/CVS-170592