Repository navigation
[TF-TRT] Enable INT8 calibration for use_implicit_batch=false when there is no dynamic shape inputs - #48244
[TF-TRT] Enable INT8 calibration for use_implicit_batch=false when there is no dynamic shape inputs#48244copybara-service[bot] merged 1 commit into
Conversation
|
|
||
| if (has_dynamic_shape_input_ && !use_implicit_batch_) { | ||
| OP_REQUIRES(context, !calibration_mode_, | ||
| errors::InvalidArgument( | ||
| "Dynamic shape mode does not support calibration")); | ||
| } |
There was a problem hiding this comment.
This change doesn't seem to match the description of this PR "Re-enable INT8 calibration with implicit batch mode".
IIUC, the PR description should be "Enable INT8 calibration for use_implicit_batch=false when there is no dynamic shape inputs".
And the detail description of the PR is "Previously we disable INT8 calibration when use_implicit_batch=false. This PR partially enable INT8 calibration for use_implicit_batch=false, that is, when the engine has only static input shapes."
|
@DEKHTIARJonathan Can you please check @bixia1s comments and keep us posted ? Thanks! |
|
@DEKHTIARJonathan Any update on this PR? Please. Thanks! |
|
Still under work, please do not close |
DEKHTIARJonathan
changed the title
|
@bixia1 : sorry I totally forgot to update this PR. |
google-ml-butler
Bot
added
kokoro:force-run
gbaned
removed
ready to pull
google-ml-butler
Bot
added
kokoro:force-run
@tfeher @bixia1
Previously we disable INT8 calibration when use_implicit_batch=false. This PR partially enable INT8 calibration for use_implicit_batch=false, that is, when the engine has only static input shapes.