shingjan opened a new issue, #12995:
URL: https://github.com/apache/tvm/issues/12995
The following TIR generated from the `executor_codegen_` of relay is not
valid and therefore shouldn't be accepted as a valid input to TVM's relay
executor.
```
@tvm.script.ir_module
class Module:
@T.prim_func
def tvmgen_default_fused_embedding_bag(p0: T.Buffer[13, "int32"], p1:
T.Buffer[13, "int32"], p2: T.Buffer[13, "int32"], p3: T.Buffer[13, "int32"],
T_embedding_bag: T.Buffer[156, "int32"]) -> None:
# function attr dict
T.func_attr({"hash": "36740c103ca5395c", "target":
T.target({"num-cores":12, "kind":"llvm", "host":T.target({"kind":"llvm",
"tag":"", "keys":["cpu"], "num-cores":12}), "tag":"", "keys":["cpu"]}),
"tir.noalias": True, "global_symbol": "tvmgen_default_fused_embedding_bag",
"from_legacy_te_schedule": True})
cse_var_1 = T.var("int32")
T.preflattened_buffer(p0, [13], dtype="int32", data=p0.data)
T.preflattened_buffer(p1, [1, 13], dtype="int32", data=p1.data)
T.preflattened_buffer(p2, [13], dtype="int32", data=p2.data)
T.preflattened_buffer(p3, [13], dtype="int32", data=p3.data)
T.preflattened_buffer(T_embedding_bag, [12, 13], dtype="int32",
data=T_embedding_bag.data)
# body
for ax0 in T.parallel(12):
T_embedding_bag[ax0 * 13:ax0 * 13 + 13] = T.broadcast(0, 13)
for k in T.serial(T.let(cse_var_1, ax0 + 1, T.if_then_else(12 ==
cse_var_1, 13, p2[cse_var_1], dtype="int32") - p2[ax0])):
cse_var_2: T.int32 = ax0 * 13
T_embedding_bag[cse_var_2:cse_var_2 + 13] =
T_embedding_bag[cse_var_2:cse_var_2 + 13] + T.if_then_else(p0[k] != -1,
p1[p0[k] * 13:p0[k] * 13 + 13] * T.broadcast(p3[k], 13), T.broadcast(0, 13),
dtype="int32x13")
```
### Expected behavior
The build should fail.
### Actual behavior
The build succeeded and could introduce unexpected behavior.
### Environment
#12993
### Steps to reproduce
```
import tvm
from tvm import relay
indices = relay.var("indices", relay.TensorType((13,), "int32"))
weights = relay.var("weights", relay.TensorType((1, 13), "int32"))
offsets = relay.var("offsets", relay.TensorType((13,), "int32"))
per_sample_weights = relay.var("per_sample_weights", relay.TensorType((13,),
"int32"))
res = relay.op.transform.embedding_bag(
indices=indices,
weights=weights,
offsets=offsets,
mode=0,
padding_idx=-1,
per_sample_weights=per_sample_weights,
include_last_offset=True,
)
func = relay.Function([indices, weights, offsets, per_sample_weights], res)
mod = tvm.IRModule.from_expr(func)
f = relay.build(mod, tvm.target.Target("llvm --num-cores=12"))
```
### Triage
* needs-triage
--
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]
