Skip to content

[Draft] Qualcomm AI Engine Direct - Unexpected graph for mutable buffer in Quantization #4627

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

Closed
shewu-quic wants to merge 1 commit into from
Closed

[Draft] Qualcomm AI Engine Direct - Unexpected graph for mutable buffer in Quantization #4627

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
3 changes: 2 additions & 1 deletion backends/qualcomm/tests/models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,11 +459,12 @@ def __init__(self):
self.register_buffer(
"k_cache",
torch.zeros((1, 1024, 12, 64), dtype=torch.float32),
persistent=False,
)

def forward(self, input_pos, k_val):
k_out = torch.ops.aten.index_put_(self.k_cache, [None, input_pos], k_val)
return k_out
return k_out + k_out


class LayerNorm(torch.nn.Module):
Expand Down
9 changes: 7 additions & 2 deletions backends/qualcomm/tests/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -293,7 +293,8 @@ def lower_module_and_test_output(

# this is needed for the ETRecord as lowering modifies the graph in-place
edge_copy = copy.deepcopy(delegated_program)

from executorch.backends.qualcomm.utils.utils import draw_graph
draw_graph("before_lower",".", delegated_program.exported_program.graph_module)
delegated_program.exported_program = to_backend(
delegated_program.exported_program, qnn_partitioner
)
Expand Down Expand Up @@ -342,7 +343,10 @@ def get_qdq_module(
custom_quant_annotations: Tuple[Callable] = (),
quant_dtype: QuantDtype = QuantDtype.use_8a8w,
) -> torch.fx.GraphModule:
# New advice Api
m = torch.export.export(module, inputs).module()
# Deprecated Api
# m = torch._export.capture_pre_autograd_graph(module, inputs)

quantizer = QnnQuantizer()
quantizer.add_custom_quant_annotations(custom_quant_annotations)
Expand All @@ -363,7 +367,8 @@ def get_qdq_module(

prepared = prepare_pt2e(m, quantizer)
prepared(*inputs)
quantized_module = convert_pt2e(prepared)
# Whether fold quantized or not
quantized_module = convert_pt2e(prepared, fold_quantize=True)
nodes = {node.target for node in quantized_module.graph.nodes}
q_and_dq = {
torch.ops.quantized_decomposed.quantize_per_tensor.default,
Expand Down