-
Notifications
You must be signed in to change notification settings - Fork 2.8k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
23 changed files
with
333 additions
and
78 deletions.
There are no files selected for viewing
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
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
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
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
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
53 changes: 53 additions & 0 deletions
53
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.cc
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,53 @@ | ||
|
|
||
| #include "contrib_ops/cuda/bert/attention_kernel_options.h" | ||
| #include "contrib_ops/cpu/bert/attention_common.h" | ||
| #include "core/platform/env_var_utils.h" | ||
|
|
||
| namespace onnxruntime { | ||
| namespace contrib { | ||
| namespace cuda { | ||
|
|
||
| // Initialize the singleton instance | ||
| AttentionKernelOptions AttentionKernelOptions::instance; | ||
|
|
||
| void AttentionKernelOptions::Initialize(int value) { | ||
| if (value > 0) { | ||
| use_flash_attention_ = (value & static_cast<int>(AttentionBackend::FLASH_ATTENTION)) > 0; | ||
| use_efficient_attention_ = (value & static_cast<int>(AttentionBackend::EFFICIENT_ATTENTION)) > 0; | ||
| use_trt_fused_attention_ = (value & static_cast<int>(AttentionBackend::TRT_FUSED_ATTENTION)) > 0; | ||
| use_unfused_ = (value & static_cast<int>(AttentionBackend::MATH)) > 0; | ||
| use_trt_flash_attention_ = (value & static_cast<int>(AttentionBackend::TRT_FLASH_ATTENTION)) > 0; | ||
| use_trt_cross_attention_ = (value & static_cast<int>(AttentionBackend::TRT_CROSS_ATTENTION)) > 0; | ||
| use_trt_causal_attention_ = (value & static_cast<int>(AttentionBackend::TRT_CAUSAL_ATTENTION)) > 0; | ||
| } else { | ||
| use_flash_attention_ = !ParseEnvironmentVariableWithDefault<bool>(attention::kDisableFlashAttention, false); | ||
| use_efficient_attention_ = !ParseEnvironmentVariableWithDefault<bool>(attention::kDisableMemoryEfficientAttention, false); | ||
| use_trt_fused_attention_ = !ParseEnvironmentVariableWithDefault<bool>(attention::kDisableFusedSelfAttention, false); | ||
| use_unfused_ = true; | ||
| use_trt_flash_attention_ = !ParseEnvironmentVariableWithDefault<bool>(attention::kDisableTrtFlashAttention, false); | ||
| use_trt_cross_attention_ = !ParseEnvironmentVariableWithDefault<bool>(attention::kDisableFusedCrossAttention, false); | ||
| use_trt_causal_attention_ = ParseEnvironmentVariableWithDefault<bool>(attention::kEnableFusedCausalAttention, false); | ||
| } | ||
|
|
||
| min_seq_len_for_flash_attention_packed_qkv_ = ParseEnvironmentVariableWithDefault<int>( | ||
| attention::kMinSeqLenForFlashAttentionPackedQKV, | ||
| attention::kDefaultMinSeqLenForFlashAttentionPackedQKV); | ||
|
|
||
| min_seq_len_for_efficient_attention_fp32_ = ParseEnvironmentVariableWithDefault<int>( | ||
| attention::kMinSeqLenForEfficientAttentionFp32, | ||
| attention::kDefaultMinSeqLenForEfficientAttentionFp32); | ||
|
|
||
| initialized_ = true; | ||
| } | ||
|
|
||
| const AttentionKernelOptions* AttentionKernelOptions::GetInstance(int sdpa_kernel, bool force_init) { | ||
| if (force_init || !instance.initialized_) { | ||
| instance.Initialize(sdpa_kernel); | ||
| } | ||
|
|
||
| return &instance; | ||
| } | ||
|
|
||
| } // namespace cuda | ||
| } // namespace contrib | ||
| } // namespace onnxruntime |
59 changes: 59 additions & 0 deletions
59
onnxruntime/contrib_ops/cuda/bert/attention_kernel_options.h
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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,59 @@ | ||
|
|
||
| #pragma once | ||
|
|
||
| namespace onnxruntime { | ||
| namespace contrib { | ||
| namespace cuda { | ||
|
|
||
| enum class AttentionBackend : int { | ||
| FLASH_ATTENTION = 1, | ||
| EFFICIENT_ATTENTION = 2, | ||
| TRT_FUSED_ATTENTION = 4, | ||
| MATH = 8, // unfused | ||
|
|
||
| // The following kernels might be deprected in the future. | ||
| TRT_FLASH_ATTENTION = 16, | ||
| TRT_CROSS_ATTENTION = 32, | ||
| TRT_CAUSAL_ATTENTION = 64, | ||
| }; | ||
|
|
||
| class AttentionKernelOptions { | ||
| public: | ||
| static const AttentionKernelOptions* GetInstance(int sdpa_kernel, bool force_init); | ||
|
|
||
| bool UseFlashAttention() const { return use_flash_attention_; } | ||
| bool UseEfficientAttention() const { return use_efficient_attention_; } | ||
| bool UseTrtFusedAttention() const { return use_trt_fused_attention_; } | ||
| bool UseUnfusedAttention() const { return use_unfused_; } | ||
| bool UseTrtFlashAttention() const { return use_trt_flash_attention_; } | ||
| bool UseTrtCrossAttention() const { return use_trt_cross_attention_; } | ||
| bool UseTrtCausalAttention() const { return use_trt_causal_attention_; } | ||
|
|
||
| int MinSeqLenForFlashAttentionPackedQkv() const { return min_seq_len_for_flash_attention_packed_qkv_; } | ||
| int MinSeqLenForEfficientAttentionFp32() const { return min_seq_len_for_efficient_attention_fp32_; } | ||
|
|
||
| protected: | ||
| void Initialize(int value); | ||
|
|
||
| private: | ||
| bool use_flash_attention_{true}; | ||
| bool use_efficient_attention_{true}; | ||
| bool use_trt_fused_attention_{true}; | ||
| bool use_unfused_{true}; | ||
| bool use_trt_flash_attention_{true}; | ||
| bool use_trt_cross_attention_{true}; | ||
|
|
||
| // Causal attention is disabled by default in #14732. | ||
| bool use_trt_causal_attention_{false}; | ||
|
|
||
| int min_seq_len_for_flash_attention_packed_qkv_{0}; | ||
|
|
||
| int min_seq_len_for_efficient_attention_fp32_{0}; | ||
|
|
||
| bool initialized_{false}; | ||
| static AttentionKernelOptions instance; | ||
| }; | ||
|
|
||
| } // namespace cuda | ||
| } // namespace contrib | ||
| } // namespace onnxruntime |
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
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
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
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
Oops, something went wrong.