Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-02-05 16:10:42 +01:00
Code Issues Packages Projects Releases Wiki Activity

HIP: add GGML_CUDA_CC_IS_* for amd familys as increasing cc archtectures for amd gpus are not supersets of eatch other (#11601)

Browse Source 
This fixes a bug where RDNA1 gpus other than gfx1010 where not handled correctly
This commit is contained in:
uvos 2025-02-02 22:08:05 +01:00 committed by GitHub
parent 90f9b88afb
commit 4d0598e144
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 11 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-cuda/common.cuh
View File

@ -61,6 +61,13 @@
#define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
#define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
#define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
#define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3)
#define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA)
#define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
#define GGML_CUDA_CC_QY1 210
#define GGML_CUDA_CC_QY2 220

4
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -1205,7 +1205,7 @@ static void ggml_cuda_op_mul_mat_cublas(
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
if (compute_capability == GGML_CUDA_CC_CDNA) {
if (GGML_CUDA_CC_IS_CDNA(compute_capability)) {
const float alpha = 1.0f;
const float beta = 0.0f;
CUBLAS_CHECK(
@ -1750,7 +1750,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
beta = &beta_f32;
}
if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_CC_CDNA) {
if (GGML_CUDA_CC_IS_CDNA(ggml_cuda_info().devices[ctx.device].cc)) {
cu_compute_type = CUBLAS_COMPUTE_32F;
alpha = &alpha_f32;
beta = &beta_f32;

2
ggml/src/ggml-cuda/mmq.cu
View File

@ -148,5 +148,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}
return (cc < GGML_CUDA_CC_RDNA3 && cc != GGML_CUDA_CC_CDNA && cc != GGML_CUDA_CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc) && !GGML_CUDA_CC_IS_GCN(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}

2
ggml/src/ggml-cuda/mmq.cuh
View File

@ -120,7 +120,7 @@ static constexpr __device__ int get_mmq_x_max_device() {
}
static constexpr int get_mmq_y_host(const int cc) {
return cc >= GGML_CUDA_CC_OFFSET_AMD ? (cc == GGML_CUDA_CC_RDNA1 ? 64 : 128) : (cc >= GGML_CUDA_CC_VOLTA ? 128 : 64);
return cc >= GGML_CUDA_CC_OFFSET_AMD ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) : (cc >= GGML_CUDA_CC_VOLTA ? 128 : 64);
}
static constexpr __device__ int get_mmq_y_device() {