Skip to content

GitLab

Commit 5e22689a authored by Vijay Korthikanti's avatar Vijay Korthikanti
Browse files

fused kernels simd fix for short sequences

parent e269e200
Hide whitespace changes
Inline Side-by-side
Showing with 2 additions and 2 deletions
megatron/fused_kernels/scaled_masked_softmax.h
View file @ 5e22689a
......@@ -111,7 +111,7 @@ __global__ void scaled_masked_softmax_warp_forward(
constexpr int WARP_SIZE = (next_power_of_two < C10_WARP_SIZE) ? next_power_of_two : C10_WARP_SIZE;
constexpr int WARP_ITERATIONS = next_power_of_two / WARP_SIZE;
constexpr int WARP_BATCH = (next_power_of_two <= 128) ? 2 : 1;
constexpr int ELEMENTS_PER_LDG_STG = 4;
constexpr int ELEMENTS_PER_LDG_STG = (WARP_ITERATIONS < 4) ? 1 : 4;
// blockDim/threadIdx = (WARP_SIZE, WARPS_PER_BLOCK, )
// gridDim/blockIdx = (seq_len, attn_heads, batches)
......@@ -230,7 +230,7 @@ __global__ void scaled_masked_softmax_warp_backward(
constexpr int WARP_SIZE = (next_power_of_two < C10_WARP_SIZE) ? next_power_of_two : C10_WARP_SIZE;
constexpr int WARP_ITERATIONS = next_power_of_two / WARP_SIZE;
constexpr int WARP_BATCH = (next_power_of_two <= 128) ? 2 : 1;
constexpr int ELEMENTS_PER_LDG_STG = 4;
constexpr int ELEMENTS_PER_LDG_STG = (WARP_ITERATIONS < 4) ? 1 : 4;
// blockDim/threadIdx = (WARP_SIZE, WARPS_PER_BLOCK, )
// gridDim/blockIdx = (seq_len, attn_heads, batches)
......
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