Ubospica commented on code in PR #16039:
URL: https://github.com/apache/tvm/pull/16039#discussion_r1384207525
##########
src/target/target_kind.cc:
##########
@@ -334,6 +334,7 @@ TVM_REGISTER_TARGET_KIND("cuda", kDLCUDA)
.add_attr_option<Integer>("max_threads_per_block")
.add_attr_option<Integer>("thread_warp_size", Integer(32))
.add_attr_option<Integer>("registers_per_block")
+ .add_attr_option<Integer>("l2_cache_size_bytes")
Review Comment:
Yeah I agree with that. It's only I am following the previously defined(but
not implemented) attribute in target, see
https://github.com/apache/tvm/blob/main/python/tvm/target/target.py#L272
And the same attribute is defined and implemented for devices. See
https://github.com/apache/tvm/blob/main/python/tvm/_ffi/runtime_ctypes.py#L492
The difference is that for targets, we will hardcoded this attibute (b/c we
may build a model on devices different from the provided target), and for
devices, we will automatically detect it from real devices.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
