Quantize slice op #37630
Quantize slice op #37630
Conversation
| Thanks for your contribution! |
| @lidanqing-intel @sfraczek please review |
| Hi, @baoachun Do you think this PR could be approved? This PR added new attributes with asExtra. Slice op is op without weights and its quantization is easier. |
XieYunshen left a comment
XieYunshen left a comment There was a problem hiding this comment.
LGTM for set_tests_properties(test_analyzer_ernie_int8 PROPERTIES TIMEOUT 120)
| } | ||
| | ||
| static const std::initializer_list<std::string> variable_names_slice = { | ||
| "a", "w1", "b", "c", "d", "e", "f"}; |
There was a problem hiding this comment.
it seems you are not using 3 variables: w1,e,f
paddle/fluid/operators/slice_op.cc Outdated
| "(string, default \"float32\"). Data type of mkldnn kernel") | ||
| .SetDefault("float32") | ||
| .InEnum({"float32", "bfloat16"}) | ||
| .InEnum({"float32", "bfloat16", "int8"}) |
There was a problem hiding this comment.
The ordering has been usually float32, int8, bfloat16, so what do you think about using that order?
| GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, slice_pattern); | ||
| | ||
| // skip if prev op and next op is not quantized | ||
| if (!(IsOpDequantized(prev_op)) && !(IsOpQuantized(next_op))) { |
There was a problem hiding this comment.
I think one parenthesis would be enough but it's okay to have them too if you prefer.
zuzg dismissed stale reviews from XieYunshen, lidanqing-vv, and wozna via 748a37d 
| @XieYunshen Could you please review again? |
| @Aganlengzi Hi, could you please help approve this PR, this PR does not add any new attribute, just added one possible value for existing attribute: "mkldnn_data_type". It was been approved before by Xieyunshen #37630 (review). But cause there is one more reviewe so developer updated code, so it need approval again. Later will avoid similar situations. |
| LGTM |
| @Aganlengzi Hi, sorry but this PR need XieYunshen's approval, you can see this in log. Thanks. |
| @XieYunshen |
XieYunshen left a comment
XieYunshen left a comment There was a problem hiding this comment.
LGTM for set_tests_properties(test_analyzer_ernie_int8 PROPERTIES TIMEOUT 120)
* quantize slice op * correct test * fix code formatting
6 participants
PR types
New features
PR changes
OPs
Describe
Quantization of slice operator.
In Ernie model (
test_quant2_int8_ernie_mkldnn) quantized 1 sliced op, what caused drop in latency by 0.595%.