Thrsu opened a new issue, #17493:
URL: https://github.com/apache/tvm/issues/17493
After applying `LiftTransformParams` transformation, during relax VM
transformation, particularly in the VM Shape Lowering phase, the following
error occurs:
```
File "/software/tvm/src/relax/backend/vm/vm_shape_lower.cc", line 310
InternalError: Check failed: (it != slot_map_.end()) is false: Var mis not
defined in the function but is referenced by m * n
```
The error seems to indicate an issue with variable scope, where m is used
but not recognized in the expected scope during shape transformation. The
variable m is defined within main, and its shape is referenced correctly in
tir_vars, yet it still causes a shape resolution failure.
### Steps to reproduce
```python
import tvm
from tvm import relax
from tvm.script import ir as I
from tvm.script import tir as T
from tvm.script import relax as R
@I.ir_module
class Module:
@T.prim_func(private=True)
def add(var_weight: T.handle, var_T_add: T.handle, m: T.int64, n:
T.int64):
T.func_attr({"tir.noalias": T.bool(True)})
weight = T.match_buffer(var_weight, (m * n,))
T_add = T.match_buffer(var_T_add, (m * n,))
for ax0 in range(m * n):
with T.block("T_add"):
v_ax0 = T.axis.spatial(m * n, ax0)
T.reads(weight[v_ax0])
T.writes(T_add[v_ax0])
T_add[v_ax0] = weight[v_ax0] + T.float32(1)
@R.function
def main(x: R.Tensor(("m", "n"), dtype="float32"), weight: R.Tensor(("m
* n",), dtype="float32")) -> R.Tensor(("m * n", 1, 1, 1), dtype="float32"):
m = T.int64()
n = T.int64()
R.func_attr({"num_input": 1})
cls = Module
with R.dataflow():
gv = R.call_tir(cls.add, (weight,), out_sinfo=R.Tensor((m * n,),
dtype="float32"), tir_vars=R.shape([m, n]))
R.output(gv)
return gv
mod = Module
with tvm.transform.PassContext(disabled_pass=["RemoveUnusedParameters"]):
mod = relax.transform.FuseTIR()(mod)
mod = tvm.relax.transform.LegalizeOps()(mod)
mod = relax.transform.LambdaLift()(mod)
mod = relax.transform.LiftTransformParams()(mod)
ex = relax.build(mod, target='llvm')
```
--
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]
