EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: clamp matmul and FA results to the max finite value (#15652)

Browse Source 
* vulkan: clamp matmul and FA results to the max finite value

* only clamp for fp16
This commit is contained in:
Jeff Bolz
2025-08-31 01:27:57 -05:00
committed by GitHub
parent 4d74393bcc
commit 94e82c7ead
7 changed files with 58 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
View File

@@ -349,6 +349,10 @@ void main() {
sum = coopMatMulAdd(mat_a, mat_b, sum);
block_k += BK;
}
#if defined(ACC_TYPE_MAX)
[[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
#endif
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator>(sum);
coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover4, ir * BM, BM), tensorViewTranspose);
@@ -388,6 +392,10 @@ void main() {
sum = coopMatMulAdd(mat_a, mat_b, sum);
block_k += BK;
}
#if defined(ACC_TYPE_MAX)
[[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
#endif
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(sum);
coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover2, ir * BM, BM), tensorViewTranspose);
@@ -428,6 +436,10 @@ void main() {
sum = coopMatMulAdd(mat_a, mat_b, sum);
block_k += BK;
}
#if defined(ACC_TYPE_MAX)
[[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
#endif
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(sum);
coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose);
@@ -485,6 +497,9 @@ void main() {
sum = coopMatMulAdd(mat_a, mat_b, sum);
}
}
#if defined(ACC_TYPE_MAX)
[[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
#endif
// Convert from ACC_TYPE to D_TYPE
coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> mat_d;