Quantization folding pass #7240
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/7240
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 2a03d6f with merge base 3f7eb3b ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
per
added
partner: arm
Collaborator
Author
|
digantdesai
reviewed
Dec 11, 2024
backends/arm/operators/op_max.py
Outdated
| ) | ||
|
|
||
| output.shape = tosa_shape(output.shape, output.dim_order) | ||
| min_output = tosa_graph.addIntermediate(output.shape, ts.DType.INT32) |
Contributor
There was a problem hiding this comment.
Suggested change
| min_output = tosa_graph.addIntermediate(output.shape, ts.DType.INT32) | |
| max_output = tosa_graph.addIntermediate(output.shape, ts.DType.INT32) |
backends/arm/operators/op_max.py
Outdated
Comment on lines
46
to
57
| x_scale = input_qparams[0].scale | ||
| x_zp = input_qparams[0].zp | ||
|
|
||
| y_scale = input_qparams[1].scale | ||
| y_zp = input_qparams[1].zp | ||
|
|
||
| assert ( | ||
| x_zp == y_zp | ||
| ), "Different zp for inputs, MAX should be quantized with shared quantization!" | ||
| assert ( | ||
| x_scale == y_scale | ||
| ), "Different scale for input, MAX should be quantized with shared quantization!" |
Contributor
There was a problem hiding this comment.
refactor this as a util to assert shared qconfigs across inputs?
|
|
||
| class SimpleQuantizeModel(torch.nn.Module): | ||
| def forward(self, x): | ||
| return x + x |
Contributor
There was a problem hiding this comment.
nit: may be make it slightly more complicated with >1 input tensors and >1 add-nodes? may be like max((x + x), (y + y))
Also and also chain of nodes i.e. q0->dq0->op1->q2->dq2->op2->q3-dq3 => q0->op1*->op2*->dq3
backends/arm/tosa_quant_utils.py
Outdated
| dim_order = tensor.dim_order | ||
| tensor.shape = [tensor.shape[i] for i in dim_order] | ||
|
|
||
| qargs = list(cast(dict[int, QuantArgs], node.meta["input_qparams"]).values()) |
Contributor
There was a problem hiding this comment.
Assert input_qparams in node.meta
backends/arm/tosa_quant_utils.py
Outdated
| """ | ||
| assert len(node.meta["output_qparams"]) == 1 | ||
|
|
||
| qargs_out = cast(dict[int, QuantArgs], node.meta["output_qparams"])[0] |
backends/arm/tosa_quant_utils.py
Outdated
| return rescaled_nodes, min_scale | ||
|
|
||
|
|
||
| def insert_rescale_node_back_to_int8( |
Contributor
There was a problem hiding this comment.
Suggested change
| def insert_rescale_node_back_to_int8( | |
| def insert_rescale_node_to_int8( |
Reuse the logic from the node visiting quantization handling, but replace the quantization parameter fetching from the node meta values. Signed-off-by: Per Åstrand <[email protected]> Change-Id: I9a7bbf6384284e60118756ec5661f6b11847aba7
Fold DQ/Q nodes into the target operators specified to the pass. Signed-off-by: Per Åstrand <[email protected]> Change-Id: I8a09dc0b887dd5f3915ca157f578ecf51772a1a2
Uses the fold DQ/Q pass to encapsulate the quantization information within the node. Signed-off-by: Per Åstrand <[email protected]> Change-Id: I3adbab7e2a23a0208a03bbc423b38c15221a4959
Signed-off-by: Per Åstrand <[email protected]> Change-Id: I9230209ed3d6cc0b5ec7a35512248648bb8380ee
Signed-off-by: Per Åstrand <[email protected]> Change-Id: I6154e13a5a6b75549862709d632ee6dd5c8b0e7f
Adds a helper function to retrieve QuantArgs from node.meta and cleanup the handling a bit by introducing the __eq__ operator for QuantArgs. Signed-off-by: Per Åstrand <[email protected]> Change-Id: I519a9a286a36a278f40ffb6c679192a54d9f940d
Signed-off-by: Per Åstrand <[email protected]> Change-Id: I2d133f4347d9999c770e5337162c222368c212f2
per
force-pushed
the
quantization_folding
branch
from
4a46eec to
2a03d6f
Compare
Collaborator
Author
|
pull / unittest / macos / macos-job (pull_request) failing seems to be unrelated (test_flamingo_vision_encoder) |
digantdesai
approved these changes
Dec 13, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
Adds a folding pass to fold in q and dq nodes.
Test plan
Added test for the new pass