Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[SYCL] supprt Flash Attention for fp32/fp16/Q4/Q5/Q8 (#20190)

Browse Source 
* support flash-attention for fp32/fp16/Q4/Q5/Q8

* rm warining

* update for JIT
This commit is contained in:
Neo Zhang
2026-03-08 12:00:07 +08:00
committed by GitHub
parent c5a778891b
commit 213c4a0b81
65 changed files with 20091 additions and 8593 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/softmax.cpp
+5 -5
View File
@@ -102,7 +102,7 @@ static void soft_max_f32(const float * x,
max_val = sycl::max(max_val, val);
}
// find the max value in the block
max_val = warp_reduce_max(max_val);
max_val = warp_reduce_max<WARP_SIZE>(max_val);
if (block_size > WARP_SIZE) {
if (warp_id == 0) {
@@ -116,7 +116,7 @@ static void soft_max_f32(const float * x,
item_ct1.barrier();
max_val = buf_iw[lane_id];
max_val = warp_reduce_max(max_val);
max_val = warp_reduce_max<WARP_SIZE>(max_val);
}
float tmp = 0.0f; // partial sum
@@ -133,7 +133,7 @@ static void soft_max_f32(const float * x,
vals[col] = val;
}
// find the sum of exps in the block
tmp = warp_reduce_sum(tmp);
tmp = warp_reduce_sum<WARP_SIZE>(tmp);
if (block_size > WARP_SIZE) {
item_ct1.barrier();
if (warp_id == 0) {
@@ -153,7 +153,7 @@ static void soft_max_f32(const float * x,
for (size_t i = 1; i < nreduce; i += 1) {
tmp += buf_iw[lane_id + i * WARP_SIZE];
}
tmp = warp_reduce_sum(tmp);
tmp = warp_reduce_sum<WARP_SIZE>(tmp);
}
if (sinks) {
tmp += sycl::native::exp(sinks[i02] - max_val);
@@ -191,7 +191,7 @@ static void soft_max_back_f32(const float *grad, const float *dstf, float *dst,
dgf_dot += dstf[col]*grad[col];
}
dgf_dot = warp_reduce_sum(dgf_dot);
dgf_dot = warp_reduce_sum<WARP_SIZE>(dgf_dot);
for (int col = tid; col < ncols; col += WARP_SIZE) {
dst[col] = scale * (grad[col] - dgf_dot) * dstf[col];