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

multiheadattention int8 quantization #5733

Merged
merged 8 commits into from
Oct 15, 2024
Merged

multiheadattention int8 quantization #5733

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
07b739e
mha int8
nihui Oct 14, 2024
9896fd8
add test
nihui Oct 14, 2024
5deab1a
apply code-format changes
nihui Oct 14, 2024
0f0f310
wip
nihui Oct 15, 2024
6e765a2
apply code-format changes
nihui Oct 15, 2024
8113cb8
x86 vulkan fallback
nihui Oct 15, 2024
f50f277
comment about bf16s
nihui Oct 15, 2024
85ddd8a
Merge branch 'mha-quantize-r' of github.com:nihui/ncnn into mha-quant…
nihui Oct 15, 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
5 changes: 5 additions & 0 deletions docs/developer-guide/operators.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1277,6 +1277,7 @@ y = affine(out)
| 4 | vdim | int | embed_dim | |
| 5 | attn_mask | int | 0 | |
| 6 | scale | float | 1.f / sqrt(embed_dim / num_heads) | |
| 18 | int8_scale_term | int | 0 | |

| weight | type | shape |
| ------------- | ----- | --------------------- |
Expand All @@ -1288,6 +1289,10 @@ y = affine(out)
| v_bias_data | float | [embed_dim] |
| out_weight_data| float/fp16/int8 | [qdim * embed_dim] |
| out_bias_data | float | [qdim] |
| q_weight_data_int8_scales| float | [embed_dim] |
| k_weight_data_int8_scales| float | [embed_dim] |
| v_weight_data_int8_scales| float | [embed_dim] |
| out_weight_data_int8_scales| float | [1] |

# MVN
```
Expand Down
42 changes: 37 additions & 5 deletions src/layer/arm/multiheadattention_arm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ MultiHeadAttention_arm::MultiHeadAttention_arm()
#endif
#endif // __ARM_NEON

support_bf16_storage = false;
support_bf16_storage = false;// TODO enable bf16 when gemm has proper out_elemtype support

q_gemm = 0;
k_gemm = 0;
Expand Down Expand Up @@ -76,10 +76,16 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(11, 0); // output_N1M
pd.set(12, 1); // output_elempack
pd.set(14, 0); // output_transpose
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
q_gemm->load_param(pd);
Mat weights[2];
Mat weights[3];
weights[0] = q_weight_data;
weights[1] = q_bias_data;
#if NCNN_INT8
weights[2] = q_weight_data_int8_scales;
#endif
q_gemm->load_model(ModelBinFromMatArray(weights));
q_gemm->create_pipeline(opt);

Expand All @@ -105,10 +111,16 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(11, 0); // output_N1M
pd.set(12, 1); // output_elempack
pd.set(14, 0); // output_transpose
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
k_gemm->load_param(pd);
Mat weights[2];
Mat weights[3];
weights[0] = k_weight_data;
weights[1] = k_bias_data;
#if NCNN_INT8
weights[2] = k_weight_data_int8_scales;
#endif
k_gemm->load_model(ModelBinFromMatArray(weights));
k_gemm->create_pipeline(opt);

Expand All @@ -134,10 +146,16 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(11, 0); // output_N1M
pd.set(12, 1); // output_elempack
pd.set(14, 0); // output_transpose
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
v_gemm->load_param(pd);
Mat weights[2];
Mat weights[3];
weights[0] = v_weight_data;
weights[1] = v_bias_data;
#if NCNN_INT8
weights[2] = v_weight_data_int8_scales;
#endif
v_gemm->load_model(ModelBinFromMatArray(weights));
v_gemm->create_pipeline(opt);

Expand All @@ -161,10 +179,18 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(9, embed_dim); // K = maxk*inch
pd.set(10, 4); // constant_broadcast_type_C = null
pd.set(11, 0); // output_N1M
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
o_gemm->load_param(pd);
Mat weights[2];
Mat weights[3];
weights[0] = out_weight_data;
weights[1] = out_bias_data;
#if NCNN_INT8
Mat out_weight_data_int8_scales(1);
out_weight_data_int8_scales[0] = out_weight_data_int8_scale;
weights[2] = out_weight_data_int8_scales;
#endif
o_gemm->load_model(ModelBinFromMatArray(weights));
o_gemm->create_pipeline(opt);

Expand All @@ -189,6 +215,9 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(10, attn_mask ? 3 : -1); // constant_broadcast_type_C
pd.set(11, 0); // output_N1M
pd.set(12, 1); // output_elempack
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
qk_gemm->load_param(pd);
qk_gemm->load_model(ModelBinFromMatArray(0));
Option opt1 = opt;
Expand All @@ -211,6 +240,9 @@ int MultiHeadAttention_arm::create_pipeline(const Option& _opt)
pd.set(11, 0); // output_N1M
pd.set(12, 1); // output_elempack
pd.set(14, 1); // output_transpose
#if NCNN_INT8
pd.set(18, int8_scale_term);
#endif
qkv_gemm->load_param(pd);
qkv_gemm->load_model(ModelBinFromMatArray(0));
Option opt1 = opt;
Expand Down
Loading
Loading