I guess what you are seeing are potential cases when the variable can have any
value in the range of the data type you selected and therefore for correctness
all those if statements are necessary?
Any reason you want to use te.var ? I am guessing due to some dynamic shape you
want to support,
Hi @donglinz, see the doc [Install from Source — tvm 0.8.dev0 documentation
(apache.org)](https://tvm.apache.org/docs/install/from_source.html#install-from-source)
```
To debug with IRs, set(USE_RELAY_DEBUG ON) and set environment variable
TVM_LOG_DEBUG.
export TVM_LOG_DEBUG="ir/transform.cc=1
Hi @leeexyz , so what is the equivalent config of dump_ir_pass=True in
PassContext? As tvm.build_config have been deprecated?
---
[Visit
Topic](https://discuss.tvm.apache.org/t/how-to-dump-irs-for-each-pass-when-building-a-model/454/9)
to respond.
You are receiving this because you enabl
I wrote some ops with te and topi, and found that when using te.var to
represent op input shape,it became much slower than constant shape,when with
the same schedule.
When I look into the generated cuda code, there are lots of if sentences about
the var,which make it much slower.So I wonder i
Oh, I think I see. It's an arithmetic trick to ensure that a dtype where bits<8
will still result in the nbytes-per-element to be 1 byte.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/ndarray-h-getdatasize-question/11485/2)
to respond.
You are receiving this because you enabled mail
In this implementation of `GetDataSize()`:
https://github.com/apache/tvm/blob/58716081637fdf21bdae44122086f1dbc35f7c35/include/tvm/runtime/ndarray.h#L316
there is this mysterious line:
```
size *= (arr.dtype.bits * arr.dtype.lanes + 7) / 8;
```
I was instead expecting the simpler:
```
size *=
I'm confused with the tvm pass and the schedule, can you tell me the difference
between tvm pass and the schedule?
---
[Visit
Topic](https://discuss.tvm.apache.org/t/question-on-how-to-manual-schedule-and-optimize-the-front-model/11472/5)
to respond.
You are receiving this because you en
"relay.build" will transfer relay IR to lower IR(TIR) first then codegen to
machine code, you can optimize the TIR in your tvm pass.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/question-on-how-to-manual-schedule-and-optimize-the-front-model/11472/4)
to respond.
You are receiving t
hi, I was doing research recently and found that tvm can only implement
single-card reasoning, and I wonder if tvm can implement multi-card or
heterogeneous reasoning？
---
[Visit
Topic](https://discuss.tvm.apache.org/t/tvm-neng-shixian-duo-k-tuil-ma-11-5000-translation-results-can-tvm-rea
Hello,
In TVM4J subproject there are folders linux-x86_64 and osx_x86_64 but there is
no assembly folders for arm32, aarch64, windows folders?
---
[Visit
Topic](https://discuss.tvm.apache.org/t/why-in-tvm4j-there-is-not-arm-folders/11482/1)
to respond.
You are receiving this because you
After reading the nimble paper, I would like to ask about 3.5 Symbolic Codegen:
1. How to combine autotvm and nimble? Does the model template and search space
need to be reduced by human control? How to find the top 100 schedulers?
2. Can ansor and nimble be combined? If so, how can it be realiz
11 matches
Mail list logo
