[Reopen][Inductor][CPU] Fuse SmoothQuant int8 linear pattern #142036
Closed
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/142036
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit f3d6c5a with merge base b576a8c ( BROKEN TRUNK - The following job 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. |
|
Hi @jerryzh168 CI shows green. Would you like to import it? Thanks. |
|
@jerryzh168 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
jerryzh168
approved these changes
Dec 5, 2024
|
@pytorchbot merge |
Merge failedReason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR! Details for Dev Infra teamRaised by workflow job |
Hi @jerryzh168 It failed to merge by pytorchmergebot. Could you please import it again? Thanks. |
|
Hi @jerryzh168 The failure looks irrelevant. Could you please take a look and see if it can be imported again? Thanks. |
sanchitintel
reviewed
Dec 6, 2024
|
Hi @jerryzh168 There is a dependent PR targeting 2.6 branch cut. Could you please check if this PR can be imported and merged? Thanks. |
|
@jerryzh168 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
jgong5
approved these changes
Dec 7, 2024
Hi @jerryzh168 We would like this in 2.6. Could you please merge this PR once internal checks show green? Thanks. |
|
yeah I'm trying to land this |
|
@jerryzh168 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@pytorchbot merge (Initiating merge automatically since Phabricator Diff has merged) |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
|
@jerryzh168 Thanks a lot |
pytorchmergebot
pushed a commit
that referenced
this pull request
Dec 13, 2024
<
9E88
input type="hidden" name="badge_size" value="small" autocomplete="off" data-targets="batch-deferred-content.inputs" />
#142110) ### Summary Extends #142036 for Inductor pattern-matching pattern covered for torchao API `int8_dynamic_activation_int8_weight` in the following scenario (inference-only, freezing enabled) - - int8 quantized (symmetrically) activation (per token quantized). - Statically (so, scales are also constant. But then they would have been constant even in case of dynamic quantization due to constant weights, anyway) per-channel int8 quantized (symmetrically) weights (which are also constant because freezing is enabled). The pattern that's matched is `torch._intmm` -> convert to FP32/BF16 -> [optional expand for activation scale] ->`mul` -> `mul`. We don't check if the activation is dynamically quantized or whether the weights are statically quantized, though (since the implementation won't have have any side-effects even if that wouldn't be true). In practice, it also matches the smooth-quant int8 quantized linear pattern if its output is not reshaped (if activation is 2D). ### More details oneDNN int8 matmul supports application of per-channel weight scale but not a vector activation scale, which could be applied as a post op, but is currently unsupported in ATen. Bias addition (which could be supported with an add post-op) is also unfused. The fusion pattern used in this PR is `torch._intmm` -> convert to FP32/BF16 ->`mul`, which will be replaced by oneDNN qlinear op. The speedup over eager-mode is due to 2 reasons - 1. fusion of int8xint8 -> int32 GEMM, conversion to FP32/BF16 & application of weight scale. (In case of BF16, many intermediate conversions are also avoided). 2. weight is pre-packed & cached by Inductor, so a reorder is avoided at run-time. But, in the future, the whole pattern (including application of activation scale, which would be a mul post-op) + bias could be fused if corresponding support would be enabled in ATen. ### Verification Added UT in this PR ``` python test/inductor/test_mkldnn_pattern_matcher.py -v -k test_da8w8_sym_act_sym_wgt_with_int_mm ``` #### Corresponding torchao UTs 1. int8 Smoothquant legacy API - `TORCHINDUCTOR_FREEZING=1 TORCH_COMPILE_DEBUG=1 TORCH_LOGS="+inductor" python test/integration/test_integration.py -v -k test_non_dynamically_quantizable_linear`. The difference from #139595 is that there are no reshapes of the linear output in this pattern. 2. int8 da8w8 - symmetrically quantized activation (dynamically) & statically quantized weights - ` TORCH_COMPILE_DEBUG=1 TORCH_LOGS="+inductor" TORCHINDUCTOR_FREEZING=1 python test/integration/test_integration.py -v -k test_int8_dynamic_quant_subclass_api_0_cpu` Pull Request resolved: #142110 Approved by: https://github.com/leslie-fang-intel, https://github.com/jgong5 ghstack dependencies: #142036
bluenote10
pushed a commit
to bluenote10/pytorch
that referenced
this pull request
Dec 14, 2024
…#142036) Reopen of pytorch#139595 **About the PR** In the implementation of SmoothQuant in Torchao, quantized linear is computed by `_int_mm(a, b)` + `mul(b_scale)` + `mul(a_scale)` (+ optional `add` for bias) with `reshape` and `convert_dtype` in between. This PR adds a pass to fuse the corresponding patterns: - (no bias) `reshape -> _int_mm -> convert_element_type -> (expand -> mul) -> mul -> reshape` - (with bias) `pattern_no_bias -> add -> reshape -> reshape` The patterns are replaced by `onednn.qlinear_pointwise` and `onednn.qlinear_prepack`, the latter of which is evaluated and frozen during the freezing process of Inductor. The final graph contains `onednn.qlinear_pointwise` only with packed weight constants. Note that `onednn.qlinear_pointwise` only supports a scalar activation scale, which is a limitation of oneDNN library, so in that case we set activation scale to 1 and bias to none and apply scales and add bias after `onednn.qlinear_pointwise`. **Validation results** Accuracy/perplexity is not changed with or without this fusion pass. Latency is improved by >10% with the fusion pass. Test method: - Model: EleutherAI/gpt-j-6b - Hardware: Intel(R) Xeon(R) Platinum 8490H, running on 1 socket, 60 cores - Using Intel OMP and Tcmalloc - Running [the example script of SmoothQuant in Torchao](https://github.com/pytorch/ao/blob/main/torchao/prototype/smoothquant/example.py) with `TORCHINDUCTOR_FREEZING=1 numactl -N1 python example.py -m EleutherAI/gpt-j-6b --device=cpu --quant-mode=dynamic --compile` **Test plan** ``` python test/inductor/test_mkldnn_pattern_matcher.py -k test_smooth_quant_with_int_mm ``` Differential Revision: [D66796966](https://our.internmc.facebook.com/intern/diff/D66796966) Pull Request resolved: pytorch#142036 Approved by: https://github.com/jerryzh168, https://github.com/jgong5 Co-authored-by: sanchitintel <sanchit.jain@intel.com>
bluenote10
pushed a commit
to bluenote10/pytorch
that referenced
this pull request
Dec 14, 2024
pytorch#142110) ### Summary Extends pytorch#142036 for Inductor pattern-matching pattern covered for torchao API `int8_dynamic_activation_int8_weight` in the following scenario (inference-only, freezing enabled) - - int8 quantized (symmetrically) activation (per token quantized). - Statically (so, scales are also constant. But then they would have been constant even in case of dynamic quantization due to constant weights, anyway) per-channel int8 quantized (symmetrically) weights (which are also constant because freezing is enabled). The pattern that's matched is `torch._intmm` -> convert to FP32/BF16 -> [optional expand for activation scale] ->`mul` -> `mul`. We don't check if the activation is dynamically quantized or whether the weights are statically quantized, though (since the implementation won't have have any side-effects even if that wouldn't be true). In practice, it also matches the smooth-quant int8 quantized linear pattern if its output is not reshaped (if activation is 2D). ### More details oneDNN int8 matmul supports application of per-channel weight scale but not a vector activation scale, which could be applied as a post op, but is currently unsupported in ATen. Bias addition (which could be supported with an add post-op) is also unfused. The fusion pattern used in this PR is `torch._intmm` -> convert to FP32/BF16 ->`mul`, which will be replaced by oneDNN qlinear op. The speedup over eager-mode is due to 2 reasons - 1. fusion of int8xint8 -> int32 GEMM, conversion to FP32/BF16 & application of weight scale. (In case of BF16, many intermediate conversions are also avoided). 2. weight is pre-packed & cached by Inductor, so a reorder is avoided at run-time. But, in the future, the whole pattern (including application of activation scale, which would be a mul post-op) + bias could be fused if corresponding support would be enabled in ATen. ### Verification Added UT in this PR ``` python test/inductor/test_mkldnn_pattern_matcher.py -v -k test_da8w8_sym_act_sym_wgt_with_int_mm ``` #### Corresponding torchao UTs 1. int8 Smoothquant legacy API - `TORCHINDUCTOR_FREEZING=1 TORCH_COMPILE_DEBUG=1 TORCH_LOGS="+inductor" python test/integration/test_integration.py -v -k test_non_dynamically_quantizable_linear`. The difference from pytorch#139595 is that there are no reshapes of the linear output in this pattern. 2. int8 da8w8 - symmetrically quantized activation (dynamically) & statically quantized weights - ` TORCH_COMPILE_DEBUG=1 TORCH_LOGS="+inductor" TORCHINDUCTOR_FREEZING=1 python test/integration/test_integration.py -v -k test_int8_dynamic_quant_subclass_api_0_cpu` Pull Request resolved: pytorch#142110 Approved by: https://github.com/leslie-fang-intel, https://github.com/jgong5 ghstack dependencies: pytorch#142036
Esquains
pushed a commit
to Esquains/study1
that referenced
this pull request
Dec 15, 2024
ghstack-source-id: 426074b Pull Request resolved: pytorch/pytorch#142036
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Stack from ghstack (oldest at bottom):
Reopen of #139595
About the PR
In the implementation of SmoothQuant in Torchao, quantized linear is computed by
_int_mm(a, b)+mul(b_scale)+mul(a_scale)(+ optionaladdfor bias) withreshapeandconvert_dtypein between.This PR adds a pass to fuse the corresponding patterns:
reshape -> _int_mm -> convert_element_type -> (expand -> mul) -> mul -> reshapepattern_no_bias -> add -> reshape -> reshapeThe patterns are replaced by
onednn.qlinear_pointwiseandonednn.qlinear_prepack, the latter of which is evaluated and frozen during the freezing process of Inductor. The final graph containsonednn.qlinear_pointwiseonly with packed weight constants.Note that
onednn.qlinear_pointwiseonly supports a scalar activation scale, which is a limitation of oneDNN library, so in that case we set activation scale to 1 and bias to none and apply scales and add bias afteronednn.qlinear_pointwise.Validation results
Accuracy/perplexity is not changed with or without this fusion pass.
Latency is improved by >10% with the fusion pass.
Test method:
TORCHINDUCTOR_FREEZING=1 numactl -N1 python example.py -m EleutherAI/gpt-j-6b --device=cpu --quant-mode=dynamic --compileTest plan
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov
Differential Revision: D66796966