Skip to content

GitLab

Unverified Commit abdfcd4f authored by Elvir Crnčević's avatar Elvir Crnčević Committed by GitHub
Browse files

silu-v1: Fix EPS not being used during max-reduction (#25069) 


Signed-off-by: default avatarelvircrn <elvircrn@gmail.com>
parent 4f02b77d
Hide whitespace changes
Inline Side-by-side
Showing with 1 addition and 2 deletions
csrc/quantization/activation_kernels.cu
View file @ abdfcd4f
......@@ -365,7 +365,6 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
int32_t compute_pipeline_offset_64 = 0;
for (int32_t t = n_tokens_lower; t < n_tokens_upper; ++t) {
__nv_bfloat16 y_max_bf16 = EPS;
__nv_bfloat162 results_bf162[2];
cp_async_wait<NUM_STAGES - 2>();
......@@ -405,7 +404,7 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
auto _y_max2 =
__hmax2(__habs2(results_bf162[0]), __habs2(results_bf162[1]));
y_max_bf16 = __hmax(_y_max2.x, _y_max2.y);
__nv_bfloat16 y_max_bf16 = __hmax(EPS, __hmax(_y_max2.x, _y_max2.y));
// An entire group is assigned to a single warp, so a simple warp reduce
// is used.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment