Add fuse() to remaining QuantizationPatterns (#19727)#19727
Open
ethansfng wants to merge 4 commits into
Open
Conversation
meta-cla
Bot
added
the
CLA Signed
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19727
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 2 Unrelated FailuresAs of commit 82a1dd9 with merge base ec76470 ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Contributor
|
@ethansfng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D105728177. |
This PR needs a
|
meta-codesync
Bot
changed the title
ethansfng
force-pushed
the
export-D105728177
branch
from
7eb7834 to
25e52f7
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 21, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 21, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
force-pushed
the
export-D105728177
branch
from
25e52f7 to
7c183db
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 21, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 21, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
force-pushed
the
export-D105728177
branch
from
7c183db to
008c464
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 21, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 22, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 22, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
force-pushed
the
export-D105728177
branch
from
008c464 to
e65a3a3
Compare
…19743) Summary: torchao's `convert_pt2e` adds `out_dtype` kwargs to dequant nodes for bf16 models. `cadence::dequantize_per_tensor` doesn't support this kwarg (it hardcodes float32 output), so `ReplacePT2DequantWithCadenceDequantPass` crashes when it forwards kwargs blindly to the cadence op. Strip `out_dtype` from kwargs before creating the cadence dequant node, and insert an `aten.to.dtype` cast after it to preserve the original output dtype semantics. Differential Revision: D105630451
Summary: Add infrastructure for per-pattern `fuse()` methods on Cadence `QuantizationPattern`: - Add `anchor_ops()` (default: `tuple(partition_types())`) and `fuse()` (default: `None`) to `QuantizationPattern` base class - Add shared fusion helpers: `_get_dequant`, `_find_quant_user`, `_insert_fused_op`, `_maybe_route_depthwise_conv1d`, `_fuse_conv`, `_fuse_linear`, `_fuse_matmul` - Add `QuantFusionPass` to `compiler_funcs.py` — shared executor that iterates patterns, matches `anchor_ops()`, calls `fuse()` with debug logging and dead code elimination Differential Revision: D105728137
Summary: Add `fuse()` implementations to the first batch of Cadence `QuantizationPattern` subclasses — the standard fully-quantized patterns that use the shared `_fuse_conv`, `_fuse_linear`, and `_fuse_matmul` helpers: - `AddmmPattern` — transpose weight + linear fusion - `AddPattern` — two-input quantized add - `AddReluBasePattern` — add+relu fusion with `anchor_ops()` override - `BmmPattern`, `MatmulPattern` — matmul fusion via `_fuse_matmul` - `CatPattern` — cat passthrough on quantized inputs - `Conv1dPattern`, `Conv2dPattern` — conv fusion via `_fuse_conv` with depthwise routing - `LayerNormPattern` — layer norm with default weight/bias creation - `LinearPattern` — linear fusion via `_fuse_linear` Differential Revision: D105728156
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
force-pushed
the
export-D105728177
branch
from
e65a3a3 to
82a1dd9
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 23, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
May 23, 2026
Summary: Add `fuse()` implementations to the remaining Cadence `QuantizationPattern` subclasses: - `MaxPool2dPattern`, `MaxPool2dWithoutIndicesPattern` — order-preserving pool on quantized values - `ReluBasePattern` (inherited by `ReluPattern0`/`1`) — relu with requantization - `ConvReluBasePattern` (inherited by `Conv1d`/`2dReluPattern0`/`1`) — conv+relu fusion with `anchor_ops()` override to match only the conv op - `SoftmaxPattern` — softmax with dummy mask/pos tensors and fake_mode metadata - `MixedW8A32LinearPattern` — weight-only quantized linear (no input/output quant) - `MixedW8A32ConvPattern` — weight-only quantized conv1d with NCL→NLC permutation - `MixedW8A32GruPattern` — weight-only quantized GRU with 4 dequantized params Differential Revision: D105728177
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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:
Add
fuse()implementations to the remaining CadenceQuantizationPatternsubclasses:MaxPool2dPattern,MaxPool2dWithoutIndicesPattern— order-preserving pool on quantized valuesReluBasePattern(inherited byReluPattern0/1) — relu with requantizationConvReluBasePattern(inherited byConv1d/2dReluPattern0/1) — conv+relu fusion withanchor_ops()override to match only the conv opSoftmaxPattern— softmax with dummy mask/pos tensors and fake_mode metadataMixedW8A32LinearPattern— weight-only quantized linear (no input/output quant)MixedW8A32ConvPattern— weight-only quantized conv1d with NCL→NLC permutationMixedW8A32GruPattern— weight-only quantized GRU with 4 dequantized paramsDifferential Revision: D105728177