xinyun/DeepSeek-V3
1
0
Fork 0
mirror of https://git.datalinker.icu/deepseek-ai/DeepSeek-V3.git synced 2025-12-08 20:34:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #969 from youkaichao/rmsnorm

Browse Source 
act_quant_kernel
This commit is contained in:
GeeeekExplorer 2025-08-27 17:14:24 +08:00
commit 9b4e9788e4
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
inference/kernel.py
View File

@ -23,7 +23,8 @@ def act_quant_kernel(x_ptr, y_ptr, s_ptr, BLOCK_SIZE: tl.constexpr, scale_fmt: t
pid = tl.program_id(axis=0) pid = tl.program_id(axis=0)
offs = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) offs = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
x = tl.load(x_ptr + offs).to(tl.float32) x = tl.load(x_ptr + offs).to(tl.float32)
amax = tl.max(tl.abs(x), 1e-4) amax = tl.max(tl.abs(x)) # reduction
amax = tl.maximum(amax, 1e-4) # clamp to 1e-4
s = amax / 448. s = amax / 448.
if scale_fmt == "ue8m0": if scale_fmt == "ue8m0":
exp = tl.math.ceil(tl.math.log2(s)) exp = tl.math.ceil(tl.math.log2(s))