Skip to content

Commit 8b0f487

Browse files
authored
Merge pull request #60 from junjihashimoto/feature/enable-cross-entropy-backward
Fix crossentropy_softmax_backward kernel
2 parents be0971a + 353fd1f commit 8b0f487

File tree

2 files changed

+2
-2
lines changed

2 files changed

+2
-2
lines changed

experimental/kernels/kernels.h

+1-1
Original file line numberDiff line numberDiff line change
@@ -677,7 +677,7 @@ fn main(@builtin(global_invocation_id) global_id : vec3<u32>) {
677677
for (var i : u32 = 0u; i < V; i++) {
678678
let p : {{precision}} = probs[probs_bt + i];
679679
let indicator : {{precision}} = select(0.0, 1.0, i == ix);
680-
atomicAdd(&dlogits[dlogits_bt + i], (p - indicator) * dloss);
680+
dlogits[dlogits_bt + i] += (p - indicator) * dloss;
681681
}
682682
}
683683
}

experimental/kernels/unittest_llmc/unittest_kernels.h

+1-1
Original file line numberDiff line numberDiff line change
@@ -20,7 +20,7 @@ extern "C" {
2020
#define USE_GPU_FOR_RESIDUAL_BACKWARD 1
2121
#define USE_GPU_FOR_SOFTMAX_FORWARD 1
2222
#define USE_GPU_FOR_CROSSENTROPY_FORWARD 1
23-
// #define USE_GPU_FOR_CROSSENTROPY_SOFTMAX_BACKWARD 1
23+
#define USE_GPU_FOR_CROSSENTROPY_SOFTMAX_BACKWARD 1
2424

2525

2626
#ifdef USE_GPU_FOR_ENCODER_FORWARD

0 commit comments

Comments
 (0)