Introduce quant_level into mixed precision

[yiliu30]

Type of Change

• Enhancement
• API change

conf = MixedPrecisionConfig(
    ...
+    quant_level="auto" # auto(default)/0/1
    ...
) 

Description

Usage & Behavior

0: Op-type-wise

conf = MixedPrecisionConfig(quant_level=0) # default precision is bf16

1. Trail 1: Convert ops into bf16 AMAP
2. Fallback ops into fp32 by op-type-wise

1: Op-Wise

conf = MixedPrecisionConfig(quant_level=1) # default precision is bf16

1. Trail 1: Convert ops into bf16 AMAP
2. Fallback ops into fp32 by op-wise

auto: Op-type-wise -> op-wise, by default

conf = MixedPrecisionConfig(quant_level="auto") # default precision is bf16

1. Trail 1: Convert ops into bf16 AMAP
2. Fallback ops into fp32 by op-type-wise
3. Fallback ops into fp32 by op-wise

• The docs will be updated by a follow-up PR

How has this PR been tested?

Pre-CI

Dependency Change?

None