-
Notifications
You must be signed in to change notification settings - Fork 76
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.
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
Autoquant #38
Merged
Merged
Autoquant #38
Conversation
This file contains 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
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Feb 22, 2024
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 67787a19d26071a4a64a49cc1955190256df94f5 Pull Request resolved: #38
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Feb 27, 2024
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: ac88c078c2982853312629278d272e2a11b187a2 Pull Request resolved: #38
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 2, 2024
Summary: currently issue where for multiple linear layers, get very slow dynamic quant results on layer linear layers. unclear why. Test Plan: python test/test.py -k "autoquant" <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 187.4432 0 AUTOTUNE addmm(65536x3840, 65536x1280, 1280x3840) bias_addmm 2.9764 ms 100.0% triton_mm_1 3.6858 ms 80.8% triton_mm_2 3.7502 ms 79.4% addmm 3.7887 ms 78.6% triton_mm_3 4.1547 ms 71.6% triton_mm_4 4.2022 ms 70.8% triton_mm_0 4.7970 ms 62.0% triton_mm_8 4.9596 ms 60.0% triton_mm_7 5.4343 ms 54.8% triton_mm_10 6.9352 ms 42.9% SingleProcess AUTOTUNE takes 5.6320 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f98800eb760> f(*args, **kwargs) 3.08 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.07677136734128 1311.548416 0 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 1311.548416 0 AUTOTUNE mixed_mm(65536x1280, 1280x3840) fallback_mixed_mm 2.5089 ms 100.0% triton_mm_13 6.4153 ms 39.1% triton_mm_14 6.6832 ms 37.5% triton_mm_12 7.0896 ms 35.4% triton_mm_16 7.5022 ms 33.4% triton_mm_15 7.8426 ms 32.0% triton_mm_19 9.5269 ms 26.3% triton_mm_20 11.2033 ms 22.4% triton_mm_17 13.1675 ms 19.1% triton_mm_18 13.8004 ms 18.2% SingleProcess AUTOTUNE takes 2.4977 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff12050> f(*args, **kwargs) 3.68 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986ff27b80> f(*args, **kwargs) 3.10 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.6846738075837493 3.1023880932480097 2144.447488 25 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 1280]), torch.Size([3840, 1280]), torch.Size([3840])) 2144.447488 25 AUTOTUNE int_mm(65536x1280, 1280x3840, 65536x3840) triton_mm_43 2.0319 ms 100.0% triton_mm_35 2.8135 ms 72.2% triton_mm_42 3.1552 ms 64.4% triton_mm_36 3.1754 ms 64.0% triton_mm_44 3.3460 ms 60.7% triton_mm_41 3.4036 ms 59.7% triton_mm_37 3.5030 ms 58.0% triton_mm_34 3.6553 ms 55.6% triton_mm_38 3.9232 ms 51.8% triton_mm_40 9.1934 ms 22.1% SingleProcess AUTOTUNE takes 8.1948 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f40> f(*args, **kwargs) 3.13 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f986cfd33a0> f(*args, **kwargs) 2.21 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 3.1286065466701984 2.210085652768612 2144.447488 22 <class 'torchao.quantization.autoquant.DefaultLinear'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2144.447488 22 AUTOTUNE addmm(65536x1280, 65536x3840, 3840x1280) bias_addmm 2.7966 ms 100.0% addmm 3.0447 ms 91.9% triton_mm_57 3.5612 ms 78.5% triton_mm_58 3.6919 ms 75.7% triton_mm_59 4.1908 ms 66.7% triton_mm_60 4.2350 ms 66.0% triton_mm_56 4.7210 ms 59.2% triton_mm_64 4.9001 ms 57.1% triton_mm_63 5.5218 ms 50.6% triton_mm_66 7.1417 ms 39.2% SingleProcess AUTOTUNE takes 6.3734 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9888dd2b30> f(*args, **kwargs) 3.33 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.autoquant.DefaultLinear'> 3.329739556647837 2228.913664 39 <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 39 AUTOTUNE mixed_mm(65536x3840, 3840x1280) fallback_mixed_mm 2.3987 ms 100.0% triton_mm_70 6.9153 ms 34.7% triton_mm_72 7.1634 ms 33.5% triton_mm_69 7.3164 ms 32.8% triton_mm_68 7.5070 ms 32.0% triton_mm_71 7.5631 ms 31.7% triton_mm_76 10.7759 ms 22.3% triton_mm_75 11.0692 ms 21.7% triton_mm_73 12.8898 ms 18.6% triton_mm_77 13.3715 ms 17.9% SingleProcess AUTOTUNE takes 6.2342 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9880133fd0> f(*args, **kwargs) 3.48 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f988175b610> f(*args, **kwargs) 3.22 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8WeightOnlyQuantizedLinearWeight'> 3.4762858413159847 3.2240213360637426 2228.913664 38 <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> (torch.Size([65536, 3840]), torch.Size([1280, 3840]), torch.Size([1280])) 2228.913664 38 AUTOTUNE int_mm(65536x3840, 3840x1280, 65536x1280) triton_mm_99 1.4307 ms 100.0% triton_mm_100 1.9041 ms 75.1% triton_mm_91 2.6079 ms 54.9% triton_mm_98 2.6363 ms 54.3% triton_mm_92 2.6691 ms 53.6% triton_mm_93 3.0178 ms 47.4% triton_mm_97 3.0233 ms 47.3% triton_mm_94 3.1872 ms 44.9% triton_mm_90 3.6072 ms 39.7% triton_mm_96 8.4695 ms 16.9% SingleProcess AUTOTUNE takes 8.1095 seconds <torch.utils.benchmark.utils.common.Measurement object at 0x7f9881782f80> f(*args, **kwargs) 145.38 ms 1 measurement, 20 runs , 1 thread <torch.utils.benchmark.utils.common.Measurement object at 0x7f9892843f70> f(*args, **kwargs) 143.98 ms 1 measurement, 20 runs , 1 thread <class 'torchao.quantization.subclass.Int8DynamicallyQuantizedLinearWeight'> 145.37517526187003 143.98446583654732 2230.364672 79 Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 2, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL (pytorch-labs/segment-anything-fast#114, huggingface/diffusion-fast@176e85f) Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 5, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL (pytorch-labs/segment-anything-fast#114, huggingface/diffusion-fast@176e85f) Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 398609989008c3bdf2c6e27403a3f7966883ba76 Pull Request resolved: #38
cpuhrsch
approved these changes
Mar 19, 2024
Follow up is a refactor and tutorial |
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 0dbb2ffd09a4fcce471af979b039162916591b7a Pull Request resolved: #38
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: fddbaf2c203a1745e8a84980f778c45162576cbc Pull Request resolved: #38
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: f268031a1702302cc6baa5f91328874677a0959b Pull Request resolved: #38
@HDCharles has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: Differential Revision: [D55103983](https://our.internmc.facebook.com/intern/diff/D55103983) [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 94089f74edf54f8e2122e91498b25306d322f3ab Pull Request resolved: #38
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: Differential Revision: [D55103983](https://our.internmc.facebook.com/intern/diff/D55103983) [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 94089f74edf54f8e2122e91498b25306d322f3ab Pull Request resolved: #38
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: Differential Revision: [D55103983](https://our.internmc.facebook.com/intern/diff/D55103983) [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 94089f74edf54f8e2122e91498b25306d322f3ab Pull Request resolved: #38
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: Differential Revision: [D55103983](https://our.internmc.facebook.com/intern/diff/D55103983) [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 19, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 37683856743b0c139b87b87c1b2c9acf92a9c15b Pull Request resolved: #38
@HDCharles has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: Differential Revision: [D55103983](https://our.internmc.facebook.com/intern/diff/D55103983) [ghstack-poisoned]
HDCharles
added a commit
that referenced
this pull request
Mar 25, 2024
Summary: Adding autoquantization functionality, using hte do_quant api we can test kernel speeds and pick the best quantization type (or no quantization) for each layer. Test Plan: python test/test.py -k "autoquant" also tested on SAM and SDXL pytorch-labs/segment-anything-fast#114 HDCharles/sdxl-fast@8d9942a Reviewers: Subscribers: Tasks: Tags: ghstack-source-id: 3c1199d84d316ae49d664b6a20ebed404734806e Pull Request resolved: #38
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Labels
CLA Signed
This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
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):
Summary: Adding autoquantization functionality, using hte do_quant api
we can test kernel speeds and pick the best quantization type (or no
quantization) for each layer.
Test Plan: python test/test.py -k "autoquant"
also tested on SAM and SDXL
pytorch-labs/segment-anything-fast#114
HDCharles/sdxl-fast@8d9942a
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: D55103983