Distribute and run LLMs with a single file.
C++ cross-platformggufllama-cpplocal-ailocal-inferencelocal-llmopen-source-aisingle-file-executablespeech-to-text
Add missing AT_HWCAP test to unbreak RPI4
Justine Tunneye38de1d34990
diff --git a/llamafile/sgemm.cpp b/llamafile/sgemm.cppindex 90ccfc37..12b9c383 100644--- a/llamafile/sgemm.cpp+++ b/llamafile/sgemm.cpp@@ -128,9 +128,10 @@ bool llamafile_sgemm(int m, int n, int k, const void *A, int lda, const void *B,(const block_q8_0 *)B, ldb, (float *)C, ldc, ith, nth,task);#elif defined(__aarch64__)- return llamafile_sgemm_q0q0s_dotprod(m, n, k, (const block_q8_0 *)A, lda,- (const block_q8_0 *)B, ldb, (float *)C, ldc, ith, nth,- task);+ if (hwcap & HWCAP_FPHP)+ return llamafile_sgemm_q0q0s_dotprod(m, n, k, (const block_q8_0 *)A, lda,+ (const block_q8_0 *)B, ldb, (float *)C, ldc, ith,+ nth, task);#endifreturn false;diff --git a/llamafile/sgemmer.inc b/llamafile/sgemmer.incindex 8ee7da03..79a07471 100644--- a/llamafile/sgemmer.inc+++ b/llamafile/sgemmer.inc@@ -38,19 +38,19 @@ class SGEMMER {if (m - m0 <= 0 || n - n0 <= 0)return;int mc, nc, mp, np;- if (VECTOR_REGISTERS >= 32 && m - m0 >= 5 && n - n0 >= 5) {+ if (VECTOR_REGISTERS >= 32 && n - n0 >= 5 && m - m0 >= 5) {mc = 5;nc = 5;gemm5x5(m0, m, n0, n);- } else if (m - m0 >= 3 && n - n0 >= 4) {+ } else if (n - n0 >= 4 && m - m0 >= 3) {mc = 3;nc = 4;gemm3x4(m0, m, n0, n);- } else if (m - m0 >= 1 && n - n0 >= 4) {+ } else if (n - n0 >= 4) {mc = 1;nc = 4;gemm1x4(m0, m, n0, n);- } else if (m - m0 >= 4 && n - n0 >= 1) {+ } else if (m - m0 >= 4) {mc = 4;nc = 1;gemm4x1(m0, m, n0, n);