jinhongyii commented on code in PR #16094:
URL: https://github.com/apache/tvm/pull/16094#discussion_r1391912059
##########
tests/python/relax/distributed/test_distributed_transform_propagate_sharding.py:
##########
@@ -1060,16 +1738,19 @@ def test_mlp_pipeline_parallelism():
def test_decoder_layer():
- # mod = relax.transform.LegalizeOps({"relax.reshape": lambda bb, call:
bb.normalize(call)})(LlamaAttentionLayer)
- mod = LlamaAttentionLayer
- after = relax.distributed.transform.PropagateSharding()(mod)
+ after =
relax.distributed.transform.PropagateSharding()(LlamaAttentionLayer)
assert_structural_equal(after, ShardedLlamaAttentionLayer)
-def test_decoder_layer_dynamic_shape():
- # mod = relax.transform.LegalizeOps({"relax.reshape": lambda bb, call:
bb.normalize(call)})(LlamaAttentionLayer)
- mod = LlamaAttentionLayerDynamicShape
+def test_decoder_layer_tir():
+ mod = relax.transform.LegalizeOps()(LlamaAttentionLayer)
Review Comment:
It's an interesting topic to allow the split axis to be dynamic, but
currently we just enforce that the axis has to be static for simplicity (it's
implicitly checked in the propagation algorithm).
I don't quite understand your suggestion. How do I construct
`ShardedLlamaAttentionLayerTIR` by not writing all the TIR primfuncs?
--
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]
