[Quantization] Rework quantization_config to use QuantKey and allow for activation override (#41566)
This commit is contained in:
@@ -40,13 +40,12 @@ from vllm.platforms import current_platform
|
||||
Fp8PerBlockOnlineLinearMethod,
|
||||
Fp8PerBlockOnlineMoEMethod,
|
||||
),
|
||||
# quantization='online with linear_scheme_override and
|
||||
# moe_scheme_override
|
||||
# quantization='online' with per-layer-kind overrides
|
||||
(
|
||||
"online",
|
||||
{
|
||||
"linear_scheme_override": "fp8_per_block",
|
||||
"moe_scheme_override": "fp8_per_tensor",
|
||||
"linear": "fp8_per_block",
|
||||
"moe": "fp8_per_tensor",
|
||||
},
|
||||
Fp8PerBlockOnlineLinearMethod,
|
||||
Fp8PerTensorOnlineMoEMethod,
|
||||
|
||||
Reference in New Issue
Block a user