ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

CUDA: use arch list for compatibility check (#11775)

Browse source 
* CUDA: use arch list for feature availability check

---------

Co-authored-by: Diego Devesa <slarengh@gmail.com>
This commit is contained in:
Johannes Gäßler 2025-02-11 00:17:22 +01:00 • committed by GitHub
parent 7b891bdc86
commit b9ab0a4d0b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 80 additions and 24 deletions
Download patch file Download diff file Expand all files Collapse all files

12
ggml/src/ggml-cuda/ggml-cuda.cu
View file

@ -1867,14 +1867,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
const int cc = ggml_cuda_info().devices[id].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
}
} else {
const int cc = ggml_cuda_info().devices[ctx.device].cc;
use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc);
any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
}
// debug helpers
@ -3205,8 +3205,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
return true;
}
const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
return cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
}
case GGML_OP_CROSS_ENTROPY_LOSS:
case GGML_OP_CROSS_ENTROPY_LOSS_BACK: