Skip to content

Commit

Permalink
[CUDA] Make cubins const (#23225)
Browse files Browse the repository at this point in the history 
### Description
Make arrays with cubin data const.


### Motivation and Context
Non-const arrays are put into the .data section which might cause
excessive memory usage in some scenarios. Making cubin arrays const
allows them to be put into the .rodata section.
  • Loading branch information
@mityada
mityada authored Jan 1, 2025
1 parent afd3e81 commit 0b87bcc
Show file tree
Hide file tree
Showing 194 changed files with 693 additions and 693 deletions.
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_32_sm70.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm70_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm70_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0xe9, 0x01, 0x00,
Expand Down Expand Up @@ -10481,7 +10481,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm70_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm70_cu_cubin_len = 125536;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm70_cu_cubin_len = 125536;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_32_sm75.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm75_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm75_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0xde, 0x04, 0x00,
Expand Down Expand Up @@ -26630,7 +26630,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm75_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm75_cu_cubin_len = 319328;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm75_cu_cubin_len = 319328;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_32_sm80.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm80_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm80_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x0c, 0x04, 0x00,
Expand Down Expand Up @@ -22150,7 +22150,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm80_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm80_cu_cubin_len = 265568;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm80_cu_cubin_len = 265568;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_32_sm86.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm86_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm86_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x04, 0x00,
Expand Down Expand Up @@ -22182,7 +22182,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm86_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm86_cu_cubin_len = 265952;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm86_cu_cubin_len = 265952;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_32_sm89.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm89_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm89_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0e, 0x04, 0x00,
Expand Down Expand Up @@ -22182,7 +22182,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_32_sm89_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm89_cu_cubin_len = 265952;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_32_sm89_cu_cubin_len = 265952;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_40_sm70.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm70_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm70_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xda, 0x03, 0x00,
Expand Down Expand Up @@ -21073,7 +21073,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm70_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm70_cu_cubin_len = 252640;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm70_cu_cubin_len = 252640;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_40_sm75.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm75_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm75_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xb1, 0x05, 0x00,
Expand Down Expand Up @@ -31121,7 +31121,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm75_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm75_cu_cubin_len = 373216;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm75_cu_cubin_len = 373216;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_40_sm80.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm80_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm80_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xac, 0x04, 0x00,
Expand Down Expand Up @@ -25553,7 +25553,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm80_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm80_cu_cubin_len = 306400;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm80_cu_cubin_len = 306400;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_40_sm86.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm86_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm86_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xaa, 0x04, 0x00,
Expand Down Expand Up @@ -25510,7 +25510,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm86_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm86_cu_cubin_len = 305888;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm86_cu_cubin_len = 305888;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_40_sm89.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm89_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm89_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xaa, 0x04, 0x00,
Expand Down Expand Up @@ -25510,7 +25510,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_40_sm89_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm89_cu_cubin_len = 305888;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_40_sm89_cu_cubin_len = 305888;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_64_sm70.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm70_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm70_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x54, 0x02, 0x00,
Expand Down Expand Up @@ -12763,7 +12763,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm70_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm70_cu_cubin_len = 152928;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm70_cu_cubin_len = 152928;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_64_sm75.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm75_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm75_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x1e, 0x06, 0x00,
Expand Down Expand Up @@ -33446,7 +33446,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm75_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm75_cu_cubin_len = 401120;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm75_cu_cubin_len = 401120;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_64_sm80.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm80_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm80_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xca, 0x04, 0x00,
Expand Down Expand Up @@ -26193,7 +26193,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm80_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm80_cu_cubin_len = 314080;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm80_cu_cubin_len = 314080;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_64_sm86.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm86_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm86_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0xca, 0x04, 0x00,
Expand Down Expand Up @@ -26203,7 +26203,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm86_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm86_cu_cubin_len = 314208;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm86_cu_cubin_len = 314208;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...a/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_128_64_sm89.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm89_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm89_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0xca, 0x04, 0x00,
Expand Down Expand Up @@ -26203,7 +26203,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_128_64_sm89_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm89_cu_cubin_len = 314208;
extern const unsigned int cubin_fmha_v2_fp16_Causal_128_64_sm89_cu_cubin_len = 314208;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_32_sm70.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm70_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm70_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x5b, 0x02, 0x00,
Expand Down Expand Up @@ -12913,7 +12913,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm70_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm70_cu_cubin_len = 154720;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm70_cu_cubin_len = 154720;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_32_sm75.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm75_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm75_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x51, 0x02, 0x00,
Expand Down Expand Up @@ -12699,7 +12699,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm75_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm75_cu_cubin_len = 152160;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm75_cu_cubin_len = 152160;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_32_sm80.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm80_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm80_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x00,
Expand Down Expand Up @@ -10961,7 +10961,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm80_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm80_cu_cubin_len = 131296;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm80_cu_cubin_len = 131296;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_32_sm86.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm86_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm86_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x00,
Expand Down Expand Up @@ -10982,7 +10982,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm86_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm86_cu_cubin_len = 131552;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm86_cu_cubin_len = 131552;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_32_sm89.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm89_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm89_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x02, 0x00,
Expand Down Expand Up @@ -10982,7 +10982,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_32_sm89_cu_cubin[] = {
0x00, 0x00, 0x00, 0x00, 0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x08, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm89_cu_cubin_len = 131552;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_32_sm89_cu_cubin_len = 131552;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_40_sm70.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm70_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm70_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xdb, 0x02, 0x00,
Expand Down Expand Up @@ -15633,7 +15633,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm70_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_40_sm70_cu_cubin_len = 187360;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_40_sm70_cu_cubin_len = 187360;

} // namespace cuda
} // namespace contrib
Expand Down
4 changes: 2 additions & 2 deletions ...da/bert/tensorrt_fused_multihead_attention/causal/fmha_v2_fp16_Causal_64_40_sm75.cubin.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ namespace onnxruntime {
namespace contrib {
namespace cuda {

unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm75_cu_cubin[] = {
extern const unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm75_cu_cubin[] = {
0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x33, 0x07, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0xbe, 0x00, 0x76, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xba, 0x02, 0x00,
Expand Down Expand Up @@ -14929,7 +14929,7 @@ unsigned char cubin_fmha_v2_fp16_Causal_64_40_sm75_cu_cubin[] = {
0xe0, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00
};
unsigned int cubin_fmha_v2_fp16_Causal_64_40_sm75_cu_cubin_len = 178912;
extern const unsigned int cubin_fmha_v2_fp16_Causal_64_40_sm75_cu_cubin_len = 178912;

} // namespace cuda
} // namespace contrib
Expand Down
Loading

0 comments on commit 0b87bcc

Please sign in to comment.