DeepSeek/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP synced 2025-06-26 18:28:11 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Support UE8M0 data format. (#206)

Browse Source
This commit is contained in:
Shifang Xu
2025-06-12 09:38:19 +08:00
committed by GitHub
parent 9ec061204e
commit 21efbe9b48
14 changed files with 255 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
csrc/kernels/utils.cuh
View File

@@ -401,6 +401,43 @@ __forceinline__ __device__ int get_lane_id() {
return lane_id;
}
constexpr float kFP8Margin = 1e-4;
constexpr float kFinfoAmaxE4M3 = 448.0f;
constexpr float kFinfoAmaxInvE4M3 = 1 / 448.0f;
__forceinline__ __device__ float fast_pow2(int x) {
// We can ensure `-126 <= x and x <= 127`
uint32_t bits_x = (x + 127) << 23;
return *reinterpret_cast<float*>(&bits_x);
}
__forceinline__ __device__ int fast_log2_ceil(float x) {
auto bits_x = *reinterpret_cast<uint32_t*>(&x);
auto exp_x = (bits_x >> 23) & 0xff;
auto man_bits = bits_x & ((1 << 23) - 1);
return exp_x - 127 + (man_bits != 0);
}
__forceinline__ __device__ void calculate_fp8_scales(float amax, float& scale, float& scale_inv, bool round_scale) {
if (round_scale) {
auto exp_scale_inv = fast_log2_ceil(amax * kFinfoAmaxInvE4M3);
scale = fast_pow2(-exp_scale_inv);
scale_inv = fast_pow2(exp_scale_inv);
} else {
scale_inv = amax * kFinfoAmaxInvE4M3;
scale = kFinfoAmaxE4M3 / amax;
}
}
template <bool kIsUE8M0, typename out_dtype_t = std::conditional_t<kIsUE8M0, uint8_t, float>>
__forceinline__ __device__ out_dtype_t extract_required_scale_format(float value) {
if constexpr (kIsUE8M0) {
return static_cast<uint8_t>((*reinterpret_cast<uint32_t*>(&value)) >> 23);
} else {
return value;
}
}
template <int kNumRanks>
__forceinline__ __device__ void
barrier_block(int** barrier_signal_ptrs, int rank) {