Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[FX] Dynamic Shapes Support #3225

Merged
merged 19 commits into from
Mar 5, 2025
Merged

[FX] Dynamic Shapes Support #3225

Show file tree
Hide file tree
Changes from 17 commits
Commits
Show all changes
19 commits
Select commit Hold shift + click to select a range
8cebf99
init
anzr299 Jan 29, 2025
fbd5779
fix
anzr299 Jan 29, 2025
eda0549
pre-commit fix
anzr299 Jan 29, 2025
36aca2f
Add test case to check data type on edge tensor shape
anzr299 Jan 29, 2025
3057cf2
clean
anzr299 Jan 29, 2025
e25e594
pre commit fix
anzr299 Jan 29, 2025
bdc43c0
Merge branch 'openvinotoolkit:develop' into fx/dynamic_shapes
anzr299 Jan 29, 2025
c56ff98
update test
anzr299 Feb 4, 2025
8a6a923
Merge branch 'fx/dynamic_shapes' of https://github.com/anzr299/nncf i…
anzr299 Feb 4, 2025
f77b395
pre commit fix && clean
anzr299 Feb 4, 2025
ced6be1
Comment Fix
anzr299 Feb 4, 2025
62979bf
comment fixes
anzr299 Feb 4, 2025
292c5f1
Merge branch 'openvinotoolkit:develop' into fx/dynamic_shapes
anzr299 Feb 26, 2025
da74767
update reference graphs
anzr299 Feb 26, 2025
107c711
apply suggested changes
anzr299 Feb 26, 2025
ef28db1
pre commit fix
anzr299 Feb 26, 2025
6ccbba8
update reference graph
anzr299 Feb 26, 2025
ad9f638
remove dynamic shapes extra test
anzr299 Mar 5, 2025
253d703
review changes
anzr299 Mar 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion nncf/experimental/torch/fx/nncf_graph_builder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,7 +196,9 @@ def get_edge_params(
else:
tensor = source_node.meta["val"]
if isinstance(tensor, torch.Tensor):
tensor_shape = tuple(tensor.shape)
tensor_shape = tuple(-1 if isinstance(i, torch.SymInt) else i for i in tensor.shape)
if isinstance(tensor, torch.SymInt):
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
if isinstance(tensor, torch.SymInt):
elif isinstance(tensor, torch.SymInt):

tensor_shape = (-1,)

if tensor_shape is None:
# TODO(dlyakhov): Refactor algorithms to always have knowns edges shapes.
Expand Down
1,110 changes: 1,110 additions & 0 deletions ...ta/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/mobilenet_v3_small.dot
View file
Open in desktop

Large diffs are not rendered by default.

497 changes: 497 additions & 0 deletions ...s/torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/resnet18.dot
View file
Open in desktop

Large diffs are not rendered by default.

5,924 changes: 5,924 additions & 0 deletions .../torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/swin_v2_s.dot
View file
Open in desktop

Large diffs are not rendered by default.

49 changes: 49 additions & 0 deletions ...reference_graphs/fx/dynamic_shapes/post_quantization_compressed/synthetic_transformer.dot
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,49 @@
strict digraph {
"0 wte_weight" [id=0, type=get_attr];
"1 linear_bias" [id=1, type=get_attr];
"2 lm_head_bias" [id=2, type=get_attr];
"3 input_ids" [id=3, type=input];
"4 embedding" [id=4, type=embedding];
"5 embedding_0_0_nncf_smooth_quant_0" [id=5, type=call_module];
"6 quantize_per_tensor_default" [id=6, type=quantize_per_tensor];
"7 dequantize_per_tensor_default" [id=7, type=dequantize_per_tensor];
"8 scale_updated_constant0" [id=8, type=get_attr];
"9 compressed_weight_updated_constant0" [id=9, type=get_attr];
"10 mul_tensor" [id=10, type=mul];
"11 zero_point_updated_constant0" [id=11, type=get_attr];
"12 sub_tensor" [id=12, type=sub];
"13 linear" [id=13, type=linear];
"14 linear_0_0_nncf_smooth_quant_0" [id=14, type=call_module];
"15 quantize_per_tensor_default_1" [id=15, type=quantize_per_tensor];
"16 dequantize_per_tensor_default_1" [id=16, type=dequantize_per_tensor];
"17 scale_updated_constant1" [id=17, type=get_attr];
"18 compressed_weight_updated_constant1" [id=18, type=get_attr];
"19 mul_tensor_1" [id=19, type=mul];
"20 zero_point_updated_constant1" [id=20, type=get_attr];
"21 sub_tensor_1" [id=21, type=sub];
"22 linear_1" [id=22, type=linear];
"23 output" [id=23, type=output];
"0 wte_weight" -> "4 embedding" [label="(10, 5)", style=solid];
"1 linear_bias" -> "13 linear" [label="(5,)", style=solid];
"2 lm_head_bias" -> "22 linear_1" [label="(10,)", style=solid];
"3 input_ids" -> "4 embedding" [label="(-1,)", style=solid];
"4 embedding" -> "5 embedding_0_0_nncf_smooth_quant_0" [label="(-1, 5)", style=solid];
"5 embedding_0_0_nncf_smooth_quant_0" -> "6 quantize_per_tensor_default" [label="(-1, 5)", style=solid];
"6 quantize_per_tensor_default" -> "7 dequantize_per_tensor_default" [label="(-1, 5)", style=solid];
"7 dequantize_per_tensor_default" -> "13 linear" [label="(-1, 5)", style=solid];
"8 scale_updated_constant0" -> "10 mul_tensor" [label="(5, 1)", style=solid];
"9 compressed_weight_updated_constant0" -> "10 mul_tensor" [label="(5, 5)", style=solid];
"10 mul_tensor" -> "12 sub_tensor" [label="(5, 5)", style=solid];
"11 zero_point_updated_constant0" -> "12 sub_tensor" [label="(5, 1)", style=solid];
"12 sub_tensor" -> "13 linear" [label="(5, 5)", style=solid];
"13 linear" -> "14 linear_0_0_nncf_smooth_quant_0" [label="(-1, 5)", style=solid];
"14 linear_0_0_nncf_smooth_quant_0" -> "15 quantize_per_tensor_default_1" [label="(-1, 5)", style=solid];
"15 quantize_per_tensor_default_1" -> "16 dequantize_per_tensor_default_1" [label="(-1, 5)", style=solid];
"16 dequantize_per_tensor_default_1" -> "22 linear_1" [label="(-1, 5)", style=solid];
"17 scale_updated_constant1" -> "19 mul_tensor_1" [label="(10, 1)", style=solid];
"18 compressed_weight_updated_constant1" -> "19 mul_tensor_1" [label="(10, 5)", style=solid];
"19 mul_tensor_1" -> "21 sub_tensor_1" [label="(10, 5)", style=solid];
"20 zero_point_updated_constant1" -> "21 sub_tensor_1" [label="(10, 1)", style=solid];
"21 sub_tensor_1" -> "22 linear_1" [label="(10, 5)", style=solid];
"22 linear_1" -> "23 output" [label="(-1, 10)", style=solid];
}
515 changes: 515 additions & 0 deletions tests/torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/unet.dot
View file
Open in desktop

Large diffs are not rendered by default.

2,011 changes: 2,011 additions & 0 deletions ...s/torch/data/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/vit_b_16.dot
View file
Open in desktop

Large diffs are not rendered by default.

63 changes: 63 additions & 0 deletions ...ta/reference_graphs/fx/dynamic_shapes/post_quantization_compressed/yolo11n_sdpa_block.dot
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,63 @@
strict digraph {
"0 x" [id=0, type=input];
"1 x_0_0_nncf_smooth_quant_0" [id=1, type=call_module];
"2 quantize_per_tensor_default" [id=2, type=quantize_per_tensor];
"3 dequantize_per_tensor_default" [id=3, type=dequantize_per_tensor];
"4 scale_updated_constant0" [id=4, type=get_attr];
"5 compressed_weight_updated_constant0" [id=5, type=get_attr];
"6 mul_tensor" [id=6, type=mul];
"7 zero_point_updated_constant0" [id=7, type=get_attr];
"8 sub_tensor" [id=8, type=sub];
"9 linear" [id=9, type=linear];
"10 quantize_per_tensor_default_1" [id=10, type=quantize_per_tensor];
"11 dequantize_per_tensor_default_1" [id=11, type=dequantize_per_tensor];
"12 slice_1" [id=12, type=slice];
"13 slice_2" [id=13, type=slice];
"14 slice_3" [id=14, type=slice];
"15 quantize_per_tensor_default_2" [id=15, type=quantize_per_tensor];
"16 dequantize_per_tensor_default_2" [id=16, type=dequantize_per_tensor];
"17 slice_4" [id=17, type=slice];
"18 slice_5" [id=18, type=slice];
"19 slice_6" [id=19, type=slice];
"20 slice_7" [id=20, type=slice];
"21 slice_8" [id=21, type=slice];
"22 slice_9" [id=22, type=slice];
"23 transpose" [id=23, type=transpose];
"24 matmul" [id=24, type=matmul];
"25 div_" [id=25, type=div_];
"26 softmax" [id=26, type=softmax];
"27 transpose_1" [id=27, type=transpose];
"28 matmul_1" [id=28, type=matmul];
"29 output" [id=29, type=output];
"0 x" -> "1 x_0_0_nncf_smooth_quant_0" [label="(1, -1, -1)", style=solid];
"1 x_0_0_nncf_smooth_quant_0" -> "2 quantize_per_tensor_default" [label="(1, -1, 4)", style=solid];
"2 quantize_per_tensor_default" -> "3 dequantize_per_tensor_default" [label="(1, -1, 4)", style=solid];
"3 dequantize_per_tensor_default" -> "9 linear" [label="(1, -1, 4)", style=solid];
"4 scale_updated_constant0" -> "6 mul_tensor" [label="(12, 1)", style=solid];
"5 compressed_weight_updated_constant0" -> "6 mul_tensor" [label="(12, 4)", style=solid];
"6 mul_tensor" -> "8 sub_tensor" [label="(12, 4)", style=solid];
"7 zero_point_updated_constant0" -> "8 sub_tensor" [label="(12, 1)", style=solid];
"8 sub_tensor" -> "9 linear" [label="(12, 4)", style=solid];
"9 linear" -> "10 quantize_per_tensor_default_1" [label="(1, -1, 12)", style=solid];
"9 linear" -> "15 quantize_per_tensor_default_2" [label="(1, -1, 12)", style=solid];
"9 linear" -> "20 slice_7" [label="(1, -1, 12)", style=solid];
"10 quantize_per_tensor_default_1" -> "11 dequantize_per_tensor_default_1" [label="(1, -1, 12)", style=solid];
"11 dequantize_per_tensor_default_1" -> "12 slice_1" [label="(1, -1, 12)", style=solid];
"12 slice_1" -> "13 slice_2" [label="(1, -1, 12)", style=solid];
"13 slice_2" -> "14 slice_3" [label="(1, -1, 12)", style=solid];
"14 slice_3" -> "24 matmul" [label="(1, -1, 4)", style=solid];
"15 quantize_per_tensor_default_2" -> "16 dequantize_per_tensor_default_2" [label="(1, -1, 12)", style=solid];
"16 dequantize_per_tensor_default_2" -> "17 slice_4" [label="(1, -1, 12)", style=solid];
"17 slice_4" -> "18 slice_5" [label="(1, -1, 12)", style=solid];
"18 slice_5" -> "19 slice_6" [label="(1, -1, 12)", style=solid];
"19 slice_6" -> "23 transpose" [label="(1, -1, 4)", style=solid];
"20 slice_7" -> "21 slice_8" [label="(1, -1, 12)", style=solid];
"21 slice_8" -> "22 slice_9" [label="(1, -1, 12)", style=solid];
"22 slice_9" -> "28 matmul_1" [label="(1, -1, 4)", style=solid];
"23 transpose" -> "24 matmul" [label="(1, 4, -1)", style=solid];
"24 matmul" -> "25 div_" [label="(1, -1, -1)", style=solid];
"25 div_" -> "26 softmax" [label="(1, -1, -1)", style=solid];
"26 softmax" -> "27 transpose_1" [label="(1, -1, -1)", style=solid];
"27 transpose_1" -> "28 matmul_1" [label="(1, -1, -1)", style=solid];
"28 matmul_1" -> "29 output" [label="(1, -1, 4)", style=solid];
}
Loading