[Inductor] [Quant] Re-structure Quantization testcase pattern matcher check #112570
Conversation
… check [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/112570
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (3 Unrelated Failures)As of commit ab5a900 with merge base 871e27a ( UNSTABLE - The following jobs failed but were likely due to flakiness present on trunk and has been marked as unstable:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
… check ghstack-source-id: d504af3d68890be1a3b4d7bd36378ebf7f79858a Pull Request resolved: #112570
…ern matcher check" **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
… check ghstack-source-id: 5b4413162ccf704e4b0fb7fd8eacb61c2bd09e65 Pull Request resolved: #112570
…ern matcher check" **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
… check ghstack-source-id: 21cabe3c665fa1f530b0ebabc4e19d5ba2bcb884 Pull Request resolved: #112570
…ern matcher check" **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
… check ghstack-source-id: a051f23e78aef9c3f9c1ec30b94a7f0f6101441a Pull Request resolved: #112570
|
Hi @jerryzh168 @jgong5, could you help to take a look of this PR again for simplifying the match checker in |
| counters["inductor"]["qlinear_weight_prepack_matcher_count"], 3 | ||
| ) | ||
| self.assertEqual( | ||
| counters["inductor"]["qlinear_weight_prepack_matcher_nodes"], 18 |
There was a problem hiding this comment.
Do you think it's possible to have a function to calculate the number of matcher nodes from the pattern? Then, we don't have to hard-code the numbers here.
There was a problem hiding this comment.
Yean, I think it depends on how to write a graph analysis function with specific heuristic for calculating the expected pattern kind and pattern number given a input fx graph (after quantization flow).
If we we can get that pattern kind and numbers, then we can define a pre-known table match from pattern kind to pattern matcher node number, for example
{conv2d_weight_prepack: 6, conv2d_relu_fp32_output: 2, conv2d_relu_int8_output: 7, .....}
…ern matcher check" **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
… check ghstack-source-id: ed686714eb513bc44da94faf7670066c3ca804fd Pull Request resolved: #112570
| counters["inductor"]["qconv2d_weight_prepack_matcher_count"], 1 | ||
| ) | ||
| self.assertEqual( | ||
| counters["inductor"]["qconv2d_weight_prepack_matcher_nodes"], 6 |
There was a problem hiding this comment.
do you need this? what is this used for?
There was a problem hiding this comment.
Ideally, we only need to check matcher_count. matcher_nodes is a stricter restriction to check a node number in each matcher pattern.
| # [convert_element_type_1, sub, mul_1, dequantize_per_channel, clone, convolution] | ||
| def matcher_check_fn(): | ||
| # 1. Dequant-Conv2D pattern matched in QConv2D weight prepack * 1 | ||
| # [convert_element_type_1, sub, mul_1, dequantize_per_channel, clone, convolution] |
There was a problem hiding this comment.
how do we get this list?
There was a problem hiding this comment.
By printing the m.nodes here, we can the nodes list in each matched pattern
pytorch/torch/_inductor/pattern_matcher.py
Line 1137 in 042445b
Thanks. Yean, |
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
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 |
… check (pytorch#112570) **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` Pull Request resolved: pytorch#112570 Approved by: https://github.com/jgong5, https://github.com/jerryzh168
… check (pytorch#112570) **Summary** This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary. **TestPlan** ``` python -m pytest test_mkldnn_pattern_matcher.py -k test_q ``` Pull Request resolved: pytorch#112570 Approved by: https://github.com/jgong5, https://github.com/jerryzh168
Stack from ghstack (oldest at bottom):
Summary
This Diff re-structures Quantization testcase pattern matcher check. Instead of checking all the pattern matched in the Inductor, we will only check the core pattern match count and node numbers such as: dequant promotion, QConv/Linear Unary and QConv Binary.
TestPlan
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler