Add tt-forge sweep for conv2d. #16178
Conversation
nkpatel-tt
requested review from
mywoodstock,
shwetankTT,
sankarmanoj-tt,
pavlejosipovic,
xanderchin,
jdesousa-TT and
sjameelTT
as code owners
tests/sweep_framework/sweeps/conv2d/short/conv2d_short_sweep.py
Outdated
Show resolved
Hide resolved
| @@ -1583,6 +1583,372 @@ | |||
| ], | |||
| "is_conv1d": [True], | |||
| }, | |||
| "tt-forge_sweep_conv2d": { | |||
There was a problem hiding this comment.
can we add one pytest entry to run these locally, similar to what we have for PyTorch traces
There was a problem hiding this comment.
I am not sure what you mean. could you please elaborate more?
There was a problem hiding this comment.
Some thing like
def test_conv2d_localrun(device, input_spec):
that we have for existing sweeps
There was a problem hiding this comment.
did you resolve this one?
There was a problem hiding this comment.
Yes it is already there in latest push.
| "input_specs": [ | ||
| # Contains following params | ||
| # [batch_size, output_channels, input_channels, input_height, input_width, kernel_height, kernel_width, stride_x, stride_y, pad_x, pad_y, groups, bias, dilation, datatype] | ||
| [1, 1024, 1024, 14, 14, 3, 3, 1, 1, 1, 1, 1, False, 1, int(ttnn.bfloat16)], |
There was a problem hiding this comment.
we seem to be missing info about whether tensors are tilted or not
There was a problem hiding this comment.
tensor tilted?? I did not see such info. in sheet. I mainly have derived above from attributes only. Let me know if I am missing something.
There was a problem hiding this comment.
"input_layouts": [
{
"mapping_from": "(d0, d1, d2, d3)",
"mapping_to": "(d0 * 196 + d1 * 14 + d2, d3)",
"memory_config": [
7,
32,
"tile<32x32, bf16>",
"dram"
]
},
{
"mapping_from": "(d0, d1, d2, d3)",
"mapping_to": "(d0 * 3072 + d1 * 3 + d2, d3)",
"memory_config": [
3145728,
3,
"bf16",
"system_memory"
]
},
{
"mapping_from": "(d0, d1, d2, d3)",
"mapping_to": "(d0 * 196 + d1 * 14 + d2, d3)",
"memory_config": [
7,
32,
"tile<32x32, bf16>",
"dram"
]
}
In data you have this type of info from this you can infer if tensor is tilized or not if it's on device or on host
There was a problem hiding this comment.
ahhh....it is tiled.
Since every test case was tiled in json, I thought it is ok to ignore. However, I just checked our tensors are row major, by default.
will correct it and also add one more field of memory.
There was a problem hiding this comment.
Not all inputs are flagged as tiled, please do not ignore as this is important. In general, look at the memory config section, it will be explicitly marked as tiled.
There was a problem hiding this comment.
Several point to note here.
-> Weight tensors are missing layout info. I have confired with @AleksKnezevic and @LPanosTT and they are not giving any specific layout while creating weight tensor. same is for our case.
-> @AleksKnezevic confirmed that these test cases do not have any bias tensors.
-> input and output tensors are of tiled layout. modified code to do same.
There was a problem hiding this comment.
Now we seem to be missing info if tensor is on host or on device?
There was a problem hiding this comment.
"system_memory" means host tensor for them as far as I understand
There was a problem hiding this comment.
-> As of now, in forge input and output buffers are in DRAM. I have added buffer type as DRAM or L1 for them.
-> For weight tensor, They are using default setting which is being used now for us also.(weight buffer's buffer type is ignored.). If I understand correctly, if we do not specify any memory_config, it should be of host only. Do let me know if I am missing something.
I also have asked @AleksKnezevic and @LPanosTT for steps(, how they are running) to run test case. Once I have steps, I shall try to run one test case with debug logs enabled and see If I am missing anything.
There was a problem hiding this comment.
yeah, for now it seems that all the weights are on host, but that will probably change. It's fine to keep as is for now
nkpatel-tt
force-pushed
the
nkpatel/tt-forge-sweep
branch
from
592ac8a to
957e890
Compare
|
New traces are huge (8k lines) maybe we should put them in a separate file |
nkpatel-tt
force-pushed
the
nkpatel/tt-forge-sweep
branch
from
957e890 to
fd041a2
Compare
Done, |
nkpatel-tt
force-pushed
the
nkpatel/tt-forge-sweep
branch
from
fd041a2 to
7e44e74
Compare
Signed-off-by: Nilaykumar Patel <[email protected]>
Signed-off-by: Nilaykumar Patel <[email protected]>
Signed-off-by: Nilaykumar Patel <[email protected]>
Signed-off-by: Nilaykumar Patel <[email protected]>
nkpatel-tt
force-pushed
the
nkpatel/tt-forge-sweep
branch
from
93bacbb to
e1f8727
Compare
Ticket
#16177
Problem description
tt-forge test cases are missing in conv2d sweep suite.
What's changed
Add sweep suite for tt-forge
Checklist