Allow compiling cuda without mmq and flash attention #11190
+45
−13
Conversation
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
github-actions
bot
added
Nvidia GPU
milot-mirdita
force-pushed
the
allow-compiling-cuda-without-mmq-and-flash-attention
branch
from
3a1d670 to
8dfe3d8
Compare
| @@ -149,6 +149,7 @@ set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING | |||
| "ggml: max. batch size for using peer access") | |||
| option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF) | |||
| option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF) | |||
| option(GGML_CUDA_FA "ggml: compile with FlashAttention" ON) | |||
There was a problem hiding this comment.
Suggested change
| option(GGML_CUDA_FA "ggml: compile with FlashAttention" ON) | |
| option(GGML_CUDA_FA "ggml: compile ggml FlashAttention kernels" ON) |
| file(GLOB SRCS "template-instances/fattn-vec*f16-f16.cu") | ||
| list(FILTER GGML_SOURCES_CUDA EXCLUDE REGEX ".*fattn.*") | ||
| list(FILTER GGML_HEADERS_CUDA EXCLUDE REGEX ".*fattn.*") | ||
| # message(FATAL_ERROR ${GGML_SOURCES_CUDA}) |
There was a problem hiding this comment.
Forgot to remove?
| @@ -151,6 +151,10 @@ typedef float2 dfloat2; | |||
| #define FLASH_ATTN_AVAILABLE | |||
| #endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1) | |||
|
|
|||
| #if !defined(GGML_CUDA_FA) | |||
| #undef FLASH_ATTN_AVAILABLE | |||
| #endif | |||
There was a problem hiding this comment.
Suggested change
| #endif | |
| #endif // !defined(GGML_CUDA_FA) |
| #include "ggml-cuda/fattn.cuh" | ||
| #endif |
There was a problem hiding this comment.
Suggested change
| #endif | |
| #endif // FLASH_ATTN_AVAILABLE |
| ggml_cuda_flash_attn_ext(ctx, dst); | ||
| break; | ||
| #else | ||
| return false; | ||
| #endif |
There was a problem hiding this comment.
Suggested change
| #endif | |
| #endif // FLASH_ATTN_AVAILABLE |
Comment on lines
+3
to
+9
| #ifdef GGML_CUDA_FORCE_CUBLAS | ||
| void ggml_cuda_op_mul_mat_q( | ||
| ggml_backend_cuda_context &, | ||
| const ggml_tensor *, const ggml_tensor *, ggml_tensor *, const char *, const float *, | ||
| const char *, float *, const int64_t, const int64_t, const int64_t, | ||
| const int64_t, cudaStream_t) {} | ||
| #else |
There was a problem hiding this comment.
Add GGML_ABORT("CUDA was compiled without MMQ support") to the function instead.
| @@ -2924,6 +2925,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ3_S); | |||
| extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S); | |||
| extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL); | |||
| extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); | |||
| #endif | |||
There was a problem hiding this comment.
Suggested change
| #endif | |
| #endif // !defined(GGML_CUDA_FORCE_CUBLAS) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
I have integrated the ProstT5 protein language into Foldseek. Thanks a lot for the great library! I am upstreaming a few fixes for issues I found in ggml during the integration. I hope that it's okay to push the changes here and that they get synced at some point to the main ggml repo.
This is the last patch in my patch series. Feel free to reject this one since it might be too specific.
I want to reduce CI compile times and binary sizes for the CUDA builds. My model doesn't profit a lot from flash attention and the i only use f16 weights, so I added options to disable the kernels that compile the longest and contribute most to binary size. For MMQ I reuse the FORCE_CUBLAS option, for flash attention I added a new GGML_CUDA_FA option.