【Quant】support ue8m0 for fp8_quant_blockwise #77153
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #77153 +/- ##
==========================================
Coverage ? 0.00%
==========================================
Files ? 1
Lines ? 19
Branches ? 0
==========================================
Hits ? 0
Misses ? 19
Partials ? 0 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
liuruyan
changed the title
liuruyan
dismissed stale reviews from zyfncg, SigureMo, and qingqing01
via
8196010
|
|
||
| - op: fp8_quant_blockwise | ||
| args: (Tensor x, float epsilon, bool using_1x128_vec_quant, bool input_transpose, bool output_scale_transpose, bool return_transpose_only, bool using_e5m2, bool using_pow2_scale) | ||
| args: (Tensor x, float epsilon, bool using_1x128_vec_quant, bool input_transpose, bool output_scale_transpose, bool return_transpose_only, bool using_e5m2, bool using_pow2_scale, bool using_ue8m0_scale) |
There was a problem hiding this comment.
using_pow2_scale 和 using_ue8m0_scale 之间会有影响吗？
There was a problem hiding this comment.
没有影响，using_pow2_scale代表使用2的幂次scale，但是类型仍为float32。using_ue8m0_scale代表不仅仅使用2的幂次scale，并且输出为int32(4个ue8m0)。当两个同时开启时会以using_ue8m0_scale为准。
并且单测中存在笛卡尔积测试样例。两者不会互相冲突
There was a problem hiding this comment.
作为正式api的话，这些细节需要在api文档里说明，否则使用者只能通过试运行来确定参数作用
There was a problem hiding this comment.
好的，现在是所有attr都没有注释说明，我下一个PR来在python api处为这个算子补充一下完整的注释吧。
6 participants
PR Category
Operator Mechanism
PR Types
New features
Description
为fp8_quant_blockwise升级支持ue8m0类型scale