[Refactor][ROCm] Migrate the RDNA3 W4A16 MoE to the oracle/experts pa… (#55522)

Signed-off-by: JartX <sagformas@epdcenter.es>
This commit is contained in:
JartX
2026-09-12 21:20:26 +08:00
committed by GitHub
parent 1b29c508bc
commit 986e2f870a
10 changed files with 453 additions and 377 deletions
+34
View File
@@ -93,6 +93,40 @@ def test_moe_wna16_accepts_channelwise_gptq_activation_order():
False,
"MoeWNA16 checkpoint layout",
),
(
WNA16MoEBackend.RDNA3,
AutoGPTQConfig(4, 128, False, True, False, {}, {}),
False,
False,
"compressed-tensors",
),
(
WNA16MoEBackend.RDNA3,
QuantizationArgs(
num_bits=4,
type=QuantizationType.INT,
strategy=QuantizationStrategy.GROUP,
symmetric=False,
dynamic=False,
group_size=128,
),
True,
False,
"asymmetric",
),
(
WNA16MoEBackend.RDNA3,
QuantizationArgs(
num_bits=4,
type=QuantizationType.INT,
strategy=QuantizationStrategy.CHANNEL,
symmetric=True,
dynamic=False,
),
False,
False,
"group-wise scales",
),
],
)
def test_wna16_oracle_rejects_incompatible_quant_structures(