tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-31 14:44:34 -04:00
Code Issues Packages Projects Releases Wiki Activity

musa: fix all warnings, re-enable -DLLAMA_FATAL_WARNINGS=ON in ci and update doc (#12611)

Browse Source 
* musa: fix all warnings

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* musa: enable -DLLAMA_FATAL_WARNINGS=ON in run.sh

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* musa: update ci doc (install ccache)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* fix Windows build issue

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Address review comments

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

* Address review comments

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

---------

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
This commit is contained in:
R0CKSTAR
2025-03-30 16:59:38 +08:00
committed by GitHub
parent d3f1f0acfb
commit 492d7f1ff7
20 changed files with 191 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
ggml/src/ggml-cuda/fattn-common.cuh
View File

@@ -315,14 +315,14 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
float vals[sizeof(int)] = {0.0f};
#pragma unroll
for (int l = 0; l < sizeof(int); ++l) {
for (int l = 0; l < int(sizeof(int)); ++l) {
vals[l] = scale * x[4*threadIdx.x + l];
}
float amax = fabsf(vals[0]);
float sum = vals[0];
#pragma unroll
for (int l = 1; l < sizeof(int); ++l) {
for (int l = 1; l < int(sizeof(int)); ++l) {
amax = fmaxf(amax, fabsf(vals[l]));
sum += vals[l];
}
@@ -338,7 +338,7 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
if (d != 0.0f) {
#pragma unroll
for (int l = 0; l < sizeof(int); ++l) {
for (int l = 0; l < int(sizeof(int)); ++l) {
q8[l] = roundf(vals[l] / d);
}
}
@@ -638,7 +638,7 @@ static __global__ void flash_attn_combine_results(
float VKQ_denominator = 0.0f;
for (int l = 0; l < parallel_blocks; ++l) {
const float diff = meta[l].x - kqmax;
const float KQ_max_scale = expf(diff);
float KQ_max_scale = expf(diff);
const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
*((uint32_t *) &KQ_max_scale) &= ftz_mask;
@@ -649,6 +649,7 @@ static __global__ void flash_attn_combine_results(
dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
}
[[noreturn]]
static void on_no_fattn_vec_case(const int D) {
if (D == 64) {
fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");