Skip to content
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.

Sign up for GitHub

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

GQA MLFloat16 cpu #22102

Merged
merged 10 commits into from
Sep 24, 2024
Merged

GQA MLFloat16 cpu #22102

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
b1a88d5
init
Sep 15, 2024
0b60013
update
Sep 16, 2024
a5a6ae0
eigen MLFloat16
Sep 16, 2024
bbbdfc3
restore fp32 softmax
Sep 18, 2024
4b4b17a
update
Sep 20, 2024
d3b43f4
use constexpr
Sep 20, 2024
d635782
update
Sep 20, 2024
a9a07e1
update
Sep 20, 2024
3c55948
fix tests
Sep 24, 2024
91fe0e8
docs
Sep 24, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 8 additions & 8 deletions docs/OperatorKernels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,12 +27,12 @@ Do not modify directly.*
|Affine|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(float)|
|AffineGrid|*in* theta:**T1**<br> *in* size:**T2**<br> *out* grid:**T1**|20+|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int64)|
|And|*in* A:**T**<br> *in* B:**T**<br> *out* C:**T1**|7+|**T** = tensor(bool)<br/> **T1** = tensor(bool)|
|ArgMax|*in* data:**T**<br> *out* reduced:**tensor(int64)**|13+|**T** = tensor(double), tensor(float), tensor(int32), tensor(int8), tensor(uint8)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int8), tensor(uint8)|
|||[1, 10]|**T** = tensor(float), tensor(int32), tensor(int8), tensor(uint8)|
|ArgMin|*in* data:**T**<br> *out* reduced:**tensor(int64)**|13+|**T** = tensor(double), tensor(float), tensor(int32)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(int32)|
|||[1, 10]|**T** = tensor(float), tensor(int32)|
|ArgMax|*in* data:**T**<br> *out* reduced:**tensor(int64)**|13+|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|||[1, 10]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|ArgMin|*in* data:**T**<br> *out* reduced:**tensor(int64)**|13+|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|||[11, 12]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|||[1, 10]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(int8), tensor(uint8)|
|Asin|*in* input:**T**<br> *out* output:**T**|7+|**T** = tensor(float)|
|Asinh|*in* input:**T**<br> *out* output:**T**|9+|**T** = tensor(float)|
|Atan|*in* input:**T**<br> *out* output:**T**|7+|**T** = tensor(float)|
Expand Down Expand Up @@ -482,7 +482,7 @@ Do not modify directly.*
|Gelu|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(float)|
|GreedySearch|*in* input_ids:**I**<br> *in* max_length:**I**<br> *in* min_length:**I**<br> *in* repetition_penalty:**T**<br> *in* vocab_mask:**I**<br> *in* prefix_vocab_mask:**I**<br> *in* attention_mask:**I**<br> *out* sequences:**I**|1+|**T** = tensor(float)|
|GridSample|*in* X:**T1**<br> *in* Grid:**T1**<br> *out* Y:**T2**|1+|**T1** = tensor(float)<br/> **T2** = tensor(float)|
|GroupQueryAttention|*in* query:**T**<br> *in* key:**T**<br> *in* value:**T**<br> *in* past_key:**T**<br> *in* past_value:**T**<br> *in* seqlens_k:**M**<br> *in* total_sequence_length:**M**<br> *in* cos_cache:**T**<br> *in* sin_cache:**T**<br> *out* output:**T**<br> *out* present_key:**T**<br> *out* present_value:**T**|1+|**M** = tensor(int32)<br/> **T** = tensor(float)|
|GroupQueryAttention|*in* query:**T**<br> *in* key:**T**<br> *in* value:**T**<br> *in* past_key:**T**<br> *in* past_value:**T**<br> *in* seqlens_k:**M**<br> *in* total_sequence_length:**M**<br> *in* cos_cache:**T**<br> *in* sin_cache:**T**<br> *out* output:**T**<br> *out* present_key:**T**<br> *out* present_value:**T**|1+|**M** = tensor(int32)<br/> **T** = tensor(float), tensor(float16)|
|Inverse|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(double), tensor(float), tensor(float16)|
|MatMulBnb4|*in* A:**T1**<br> *in* B:**T2**<br> *in* absmax:**T1**<br> *out* Y:**T1**|1+|**T1** = tensor(float)<br/> **T2** = tensor(uint8)|
|MatMulFpQ4|*in* A:**T1**<br> *in* B:**T2**<br> *in* B_shape:**T3**<br> *out* Y:**T1**|1+|**T1** = tensor(float)<br/> **T2** = tensor(uint8)<br/> **T3** = tensor(int64)|
Expand All @@ -508,7 +508,7 @@ Do not modify directly.*
|QuantizeLinear|*in* x:**T1**<br> *in* y_scale:**T1**<br> *in* y_zero_point:**T2**<br> *out* y:**T2**|1+|**T1** = tensor(float)<br/> **T2** = tensor(int16), tensor(int4), tensor(int8), tensor(uint16), tensor(uint4), tensor(uint8)|
|QuickGelu|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(float)|
|Range|*in* start:**T**<br> *in* limit:**T**<br> *in* delta:**T**<br> *out* Y:**T**|1+|**T** = tensor(double), tensor(float), tensor(int16), tensor(int32), tensor(int64)|
|RotaryEmbedding|*in* input:**T**<br> *in* position_ids:**M**<br> *in* cos_cache:**T**<br> *in* sin_cache:**T**<br> *out* output:**T**|1+|**M** = tensor(int64)<br/> **T** = tensor(float)|
|RotaryEmbedding|*in* input:**T**<br> *in* position_ids:**M**<br> *in* cos_cache:**T**<br> *in* sin_cache:**T**<br> *out* output:**T**|1+|**M** = tensor(int64)<br/> **T** = tensor(float), tensor(float16)|
|SampleOp|*in* X:**T**<br> *out* Y:**T**|1+|**T** = tensor(float)|
|Sampling|*in* input_ids:**I**<br> *in* max_length:**I**<br> *in* min_length:**I**<br> *in* repetition_penalty:**T**<br> *in* vocab_mask:**I**<br> *in* prefix_vocab_mask:**I**<br> *in* attention_mask:**I**<br> *in* presence_mask:**I**<br> *in* seed:**I**<br> *out* sequences:**I**<br> *out* filtered_logits:**T**|1+|**T** = tensor(float)|
|SkipLayerNormalization|*in* input:**T**<br> *in* skip:**T**<br> *in* gamma:**T**<br> *in* beta:**T**<br> *in* bias:**T**<br> *out* output:**T**<br> *out* mean:**U**<br> *out* inv_std_var:**U**<br> *out* input_skip_bias_sum:**T**|1+|**T** = tensor(double), tensor(float)|
Expand Down
20 changes: 14 additions & 6 deletions onnxruntime/contrib_ops/cpu/bert/attention_utils.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,13 +48,13 @@
constexpr size_t element_size = sizeof(T);
ProcessBroadcastSpanFuncs add_funcs{
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.ScalarInput0<T>() + per_iter_bh.EigenInput1<T>().array();
per_iter_bh.OutputEigen<float>() = per_iter_bh.ScalarInput0<float>() + per_iter_bh.EigenInput1<float>().array();
},
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.EigenInput0<T>().array() + per_iter_bh.ScalarInput1<T>();
per_iter_bh.OutputEigen<float>() = per_iter_bh.EigenInput0<float>().array() + per_iter_bh.ScalarInput1<float>();
},
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.EigenInput0<T>() + per_iter_bh.EigenInput1<T>();
per_iter_bh.OutputEigen<float>() = per_iter_bh.EigenInput0<float>() + per_iter_bh.EigenInput1<float>();
}}; // For element-wise add

// Allocate space for output of Q(BS, D) + bias(D)
Expand Down Expand Up @@ -132,13 +132,13 @@
constexpr size_t element_size = sizeof(T);
ProcessBroadcastSpanFuncs add_funcs{
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.ScalarInput0<T>() + per_iter_bh.EigenInput1<T>().array();
per_iter_bh.OutputEigen<float>() = per_iter_bh.ScalarInput0<float>() + per_iter_bh.EigenInput1<float>().array();
},
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.EigenInput0<T>().array() + per_iter_bh.ScalarInput1<T>();
per_iter_bh.OutputEigen<float>() = per_iter_bh.EigenInput0<float>().array() + per_iter_bh.ScalarInput1<float>();
},
[](BroadcastHelper& per_iter_bh) {
per_iter_bh.OutputEigen<T>() = per_iter_bh.EigenInput0<T>() + per_iter_bh.EigenInput1<T>();
per_iter_bh.OutputEigen<float>() = per_iter_bh.EigenInput0<float>() + per_iter_bh.EigenInput1<float>();
}}; // For element-wise add

// Get Q's bias from combined bias
Expand Down Expand Up @@ -219,6 +219,10 @@
int batch_size, int num_heads, int sequence_length, int head_size,
const Tensor* in, const Tensor* bias, int bias_offset, OrtValue& out);

template Status MaybeTransposeToBNSHAndAddBias<MLFloat16>(OpKernelContext* context, AllocatorPtr allocator,
int batch_size, int num_heads, int sequence_length, int head_size,

Check warning on line 223 in onnxruntime/contrib_ops/cpu/bert/attention_utils.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Lines should be <= 120 characters long  [whitespace/line_length] [2]

Raw Output:
onnxruntime/contrib_ops/cpu/bert/attention_utils.cc:223:  Lines should be <= 120 characters long  [whitespace/line_length] [2]
const Tensor* in, const Tensor* bias, int bias_offset, OrtValue& out);

Check warning on line 224 in onnxruntime/contrib_ops/cpu/bert/attention_utils.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Lines should be <= 120 characters long  [whitespace/line_length] [2]

Raw Output:
onnxruntime/contrib_ops/cpu/bert/attention_utils.cc:224:  Lines should be <= 120 characters long  [whitespace/line_length] [2]

template <typename T>
Status MaybeTransposeToBNSH(AllocatorPtr allocator,
int batch_size, int num_heads, int sequence_length, int head_size,
Expand All @@ -242,5 +246,9 @@
int batch_size, int num_heads, int sequence_length, int head_size,
const Tensor* in, OrtValue& out);

template Status MaybeTransposeToBNSH<MLFloat16>(AllocatorPtr allocator,
int batch_size, int num_heads, int sequence_length, int head_size,
const Tensor* in, OrtValue& out);

} // namespace contrib
} // namespace onnxruntime
Loading
Loading