[Quantization][Refactor][1/N] Adopt QuantKey in QuarkConfig and methods, relying on weight_quant_key, act_quant_key for quant method dispatch (#52958)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
fxmarty-amd
2026-09-01 02:55:54 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent d0e695a91b
commit 63988f3c2d
13 changed files with 1574 additions and 476 deletions
@@ -3,11 +3,32 @@
from enum import Enum
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
kMxfp4Dynamic,
kMxfp4Static,
kMxfp6E2M3Dynamic,
kMxfp6E2M3Static,
kMxfp6E3M2Dynamic,
kMxfp6E3M2Static,
)
logger = init_logger(__name__)
OCP_MX_BLOCK_SIZE = 32
_WEIGHT_QUANT_KEY_MAP: dict[str, QuantKey] = {
"mxfp4": kMxfp4Static,
"mxfp6_e3m2": kMxfp6E3M2Static,
"mxfp6_e2m3": kMxfp6E2M3Static,
}
_ACTIVATION_QUANT_KEY_MAP: dict[str, QuantKey] = {
"mxfp4": kMxfp4Dynamic,
"mxfp6_e3m2": kMxfp6E3M2Dynamic,
"mxfp6_e2m3": kMxfp6E2M3Dynamic,
}
OCP_MX_DTYPES = {
"mxfp4",
"mxfp6_e3m2",
@@ -222,6 +222,11 @@ kFp8Dynamic128Sym = QuantKey(FP8_DTYPE, kDynamic128Scale, symmetric=True)
kStatic128BlockScale = ScaleDesc(torch.float32, True, GroupShape(128, 128))
kFp8Static128BlockSym = QuantKey(FP8_DTYPE, kStatic128BlockScale, symmetric=True)
kFp8Static128BlockE8M0Sym = QuantKey(
FP8_DTYPE,
ScaleDesc(torch.float8_e8m0fnu, True, GroupShape(128, 128)),
symmetric=True,
)
kMxfp8StaticScale = ScaleDesc(torch.uint8, True, GroupShape(1, 32))
kMxfp8Static = QuantKey(FP8_DTYPE, kMxfp8StaticScale, symmetric=True)
@@ -284,6 +289,9 @@ kInt8StaticChannelSym = QuantKey(torch.int8, kStaticChannelScale, symmetric=True
kInt8DynamicTokenSym = QuantKey(torch.int8, kDynamicTokenScale, symmetric=True)
kInt8StaticTensorSym = QuantKey(torch.int8, kStaticTensorScale, symmetric=True)
kInt8DynamicTensorSym = QuantKey(torch.int8, kDynamicTensorScale, symmetric=True)
kInt8DynamicTokenAsym = QuantKey(torch.int8, kDynamicTokenScale, symmetric=False)
kInt8StaticTensorAsym = QuantKey(torch.int8, kStaticTensorScale, symmetric=False)
kInt8DynamicTensorAsym = QuantKey(torch.int8, kDynamicTensorScale, symmetric=False)
# INT4 W4A8 quantization keys