[ROCm][Kimi-K3] Add opt-in gfx942 MXFP4-to-int4 conversion (#51274)

Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com>
Signed-off-by: Hartikainen <mahartik@amd.com>
This commit is contained in:
Markus Hartikainen
2026-10-03 14:28:02 +08:00
committed by GitHub
parent ca0df1f122
commit b0e21b3083
5 changed files with 346 additions and 2 deletions
@@ -23,6 +23,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8DynamicTokenSym,
kFp8Static128BlockSym,
kFp8StaticTensorSym,
kInt4Static32,
kInt8StaticChannelSym,
kMxfp8Dynamic,
)
@@ -46,6 +47,12 @@ def test_quant_spec_string_representation_uses_quantization_name():
assert str(quant_spec(weight="mxfp4")) == "mxfp4"
def test_quant_spec_resolves_groupwise_int4_weight():
spec = QuantSpec(weight="int4_per_group_32")
assert spec.weight == kInt4Static32
assert spec.activation is None
def test_quant_spec_rejects_unknown_name():
with pytest.raises(ValueError, match="unknown quantization name"):
quant_spec(weight="not_a_real_format")