Skip to content

Commit 325390e

Browse files
Ekaterina Ignashevafacebook-github-bot
Ekaterina Ignasheva
authored and
facebook-github-bot
committed
Temporary Commit at 5/19/2025, 10:38:46 AM
Differential Revision: D74907087
1 parent d892b66 commit 325390e

File tree

1 file changed

+109
-63
lines changed

1 file changed

+109
-63
lines changed

‎backends/cadence/aot/tests/test_fusion_ops_passes.py

Lines changed: 109 additions & 63 deletions
Original file line numberDiff line numberDiff line change
@@ -13,10 +13,6 @@
1313
import executorch.backends.cadence.aot.ops_registrations # noqa
1414
import torch
1515
from executorch.backends.cadence.aot import compiler
16-
from executorch.backends.cadence.aot.compiler import (
17-
export_to_edge,
18-
quantize_and_export_to_edge,
19-
)
2016
from executorch.backends.cadence.aot.fuse_ops import (
2117
FuseFullThenReshapePass,
2218
FuseMulScalarIntoDequantPass,
@@ -336,94 +332,144 @@ def test_replace_quant_view_dequant_with_requantize(self):
336332
)
337333

338334
def test_replace_dequant_quant_with_requantize(self):
339-
class M(torch.nn.Module):
340-
def __init__(self):
341-
super().__init__()
342-
343-
def forward(self, x):
344-
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
345-
x, 1.2, 3, 0, 127, torch.int8
346-
)
347-
x = torch.permute(x, [2, 0, 1, 3])
348-
x = torch.ops.quantized_decomposed.quantize_per_tensor(
349-
x, 4.5, 6, 0, 127, torch.int8
350-
)
351-
return x
352-
353-
inputs = torch.randn(2, 12, 1, 6).to(torch.int8)
354-
model = M()
355-
graph_module = export_to_edge(model, (inputs,)).exported_program().graph_module
356-
graph_module = FuseQuantDequantToRequantizePass()(graph_module).graph_module
335+
builder = GraphBuilder()
336+
x = builder.placeholder("x", torch.randn(2, 12, 1, 6, dtype=torch.float32))
337+
dequant = builder.call_operator(
338+
op=exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
339+
args=(x, 1.2, 3, 0, 127, torch.int8),
340+
)
341+
quant = builder.call_operator(
342+
op=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
343+
args=(dequant, 4.5, 6, 0, 127, torch.int8),
344+
)
345+
builder.output(quant)
346+
graph_module = FuseQuantDequantToRequantizePass()(
347+
builder.get_graph_module()
348+
).graph_module
357349

358350
self.check_op_counts(
359351
graph_module,
360352
expected_op_counts={
361-
# Verify that dequant -> permute -> quant was replaced with permute -> requantize.
353+
# Verify that dequant -> quant was replaced with requantize.
362354
exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: 0,
363355
exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: 0,
364356
exir_ops.edge.cadence.requantize.default: 1,
365357
},
366358
)
367359

368360
def test_replace_dequant_permute_quant_with_requantize(self):
369-
class M(torch.nn.Module):
370-
def __init__(self):
371-
super().__init__()
372-
373-
def forward(self, x):
374-
x = torch.ops.quantized_decomposed.dequantize_per_tensor(
375-
x, 1.2, 3, 0, 127, torch.int8
376-
)
377-
x = torch.permute(x, [2, 0, 1, 3])
378-
x = torch.ops.quantized_decomposed.quantize_per_tensor(
379-
x, 4.5, 6, 0, 127, torch.int8
380-
)
381-
return x
382-
383-
inputs = torch.randn(2, 12, 1, 6).to(torch.int8)
384-
model = M()
385-
graph_module = export_to_edge(model, (inputs,)).exported_program().graph_module
386-
graph_module = FuseQuantDequantToRequantizePass()(graph_module).graph_module
361+
builder = GraphBuilder()
362+
x = builder.placeholder("x", torch.randn(2, 12, 1, 6, dtype=torch.float32))
363+
dequant = builder.call_operator(
364+
op=exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
365+
args=(x, 1.2, 3, 0, 127, torch.int8),
366+
)
367+
permute = builder.call_operator(
368+
op=exir_ops.edge.aten.permute_copy.default, args=(dequant, [2, 0, 1, 3])
369+
)
370+
quant = builder.call_operator(
371+
op=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
372+
args=(permute, 4.5, 6, 0, 127, torch.int8),
373+
)
374+
builder.output(quant)
375+
graph_module = FuseQuantDequantToRequantizePass()(
376+
builder.get_graph_module()
377+
).graph_module
387378

388379
self.check_op_counts(
389380
graph_module,
390381
expected_op_counts={
391382
# Verify that dequant -> permute -> quant was replaced with permute -> requantize.
392383
exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: 0,
393384
exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: 0,
385+
exir_ops.edge.aten.permute_copy.default: 1,
394386
exir_ops.edge.cadence.requantize.default: 1,
395387
},
396388
)
397389

398390
def test_remove_nop_dequant_quant(self):
399-
class M(torch.nn.Module):
400-
def __init__(self):
401-
super(M, self).__init__()
402-
self.lin1 = torch.nn.Linear(6, 12, bias=False)
403-
self.lin2 = torch.nn.Linear(12, 24, bias=False)
391+
LEADING_DIMS: Final[int] = 12
392+
IN_DIM: Final[int] = 6
393+
OUT_DIM: Final[int] = 12
404394

405-
def forward(self, x):
406-
x = self.lin1(x)
407-
# redundant dequant+quant will be created around this permute
408-
x = torch.permute(x, [0, 2, 1, 3])
409-
x = self.lin2(x)
410-
return x
411-
412-
inputs = torch.randn(2, 12, 1, 6)
413-
model = M()
414-
graph_module = (
415-
quantize_and_export_to_edge(model, (inputs,))
416-
.exported_program()
417-
.graph_module
395+
builder = GraphBuilder()
396+
x = builder.placeholder(
397+
"x", torch.randn(LEADING_DIMS, IN_DIM, dtype=torch.float32)
398+
)
399+
quant1 = builder.call_operator(
400+
op=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
401+
args=(x, 4.5, 6, 0, 127, torch.int8),
402+
)
403+
weights = builder.call_operator(
404+
op=exir_ops.edge.aten.full.default, args=([OUT_DIM, IN_DIM], 1)
405+
)
406+
bias = builder.call_operator(
407+
op=exir_ops.edge.aten.full.default, args=([OUT_DIM], 1)
408+
)
409+
weight_zero_point = builder.call_operator(
410+
op=exir_ops.edge.aten.full.default, args=([IN_DIM], 0)
411+
)
412+
out_multiplier = builder.call_operator(
413+
op=exir_ops.edge.aten.full.default, args=([OUT_DIM], 1)
414+
)
415+
out_shift = builder.call_operator(
416+
op=exir_ops.edge.aten.full.default, args=([OUT_DIM], 0)
418417
)
419-
graph_module = FuseQuantDequantToRequantizePass()(graph_module).graph_module
418+
linear1 = builder.call_operator(
419+
op=exir_ops.edge.cadence.quantized_linear.default,
420+
args=(
421+
quant1,
422+
weights,
423+
bias,
424+
0, # src_zero_point
425+
weight_zero_point,
426+
out_multiplier,
427+
out_shift,
428+
0, # out_zero_point
429+
None,
430+
),
431+
)
432+
dequant1 = builder.call_operator(
433+
op=exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
434+
args=(linear1, 1.2, 3, 0, 127, torch.int8),
435+
)
436+
permute = builder.call_operator(
437+
op=exir_ops.edge.aten.permute_copy.default, args=(dequant1, [1, 0])
438+
)
439+
quant2 = builder.call_operator(
440+
op=exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
441+
args=(permute, 4.5, 6, 0, 127, torch.int8),
442+
)
443+
linear2 = builder.call_operator(
444+
op=exir_ops.edge.cadence.quantized_linear.default,
445+
args=(
446+
quant2,
447+
weights,
448+
bias,
449+
0, # src_zero_point
450+
weight_zero_point,
451+
out_multiplier,
452+
out_shift,
453+
0, # out_zero_point
454+
None,
455+
),
456+
)
457+
dequant2 = builder.call_operator(
458+
op=exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
459+
args=(linear2, 1.2, 3, 0, 127, torch.int8),
460+
)
461+
builder.output(dequant2)
462+
graph_module = FuseQuantDequantToRequantizePass()(
463+
builder.get_graph_module()
464+
).graph_module
420465
self.check_op_counts(
421466
graph_module,
422467
expected_op_counts={
423-
# Verify that one dequant/quant pair was removed
424-
# Expect 1 quantize ops: 1 input
468+
# Verify that one dequant/quant pair was removed from chain:
469+
# quant->linear->dequant->permute->quant->linear->dequant
470+
# gets converted to:
471+
# quant->linear->permute->linear->dequant
425472
exir_ops.edge.quantized_decomposed.quantize_per_tensor.default: 1,
426-
# Expect 1 dequant op at the end (output of second linear)
427473
exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default: 1,
428474
},
429475
)

0 commit comments

Comments
 (0)