Re: [PR] [Relax][Frontent] "tensor_ir_inplace" op [tvm]

["csullivan (via GitHub)" <gi...@apache.org>]

csullivan commented on code in PR #16498:
URL: https://github.com/apache/tvm/pull/16498#discussion_r1516748935


##########
tests/python/relax/test_frontend_nn_op.py:
##########
@@ -619,6 +619,111 @@ def test(qkv: R.Tensor((1, 1, 24, 16), dtype="float16"), offset: R.Shape(["offse
     tvm.ir.assert_structural_equal(irmodule, Expected)
 
 
+def test_tensor_ir_inplace_op():
+    hidden_size = 4096
+    dtype = "float16"
+
+    @T.prim_func
+    def inplace_take(
+        var_weight: T.handle, var_pos: T.handle, var_embeddings: T.handle, offset: T.int64
+    ):
+        T.func_attr({"tir.noalias": T.bool(True)})
+        vocab_size = T.int64()
+        weight = T.match_buffer(var_weight, (vocab_size, hidden_size), dtype)
+        seq_len = T.int64()
+        total_seq_len = T.int64()
+        pos = T.match_buffer(var_pos, (seq_len,), "int32")
+        embeddings = T.match_buffer(var_embeddings, (total_seq_len, hidden_size), dtype)
+        for ax0, ax1 in T.grid(seq_len, hidden_size):
+            with T.block("T_take"):
+                v0, v1 = T.axis.remap("SS", [ax0, ax1])
+                T.reads(weight[pos[v0], v1], pos[v0])
+                T.writes(embeddings[v0, v1])
+                embeddings[v0 + offset, v1] = weight[pos[v0], v1]
+
+    class Model(Module):
+        def test(
+            self, embedding_table: Tensor, input_ids: Tensor, embedding_dst: Tensor, offset: int
+        ):
+            tensor_expr_op_out = op.tensor_ir_op(

Review Comment:
   Did you mean to call `op.tensor_ir_inplace_op` here? It doesn't look like you are testing the new `nn.op.tensor_ir_inplace_op` added in this PR. 



-- 
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: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org