[ROCm][Quantization] add quark w4a8 mxfp4_fp8 for LinearLayer (#35316)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
This commit is contained in:
Divakar Verma
2026-03-13 19:44:24 +00:00
committed by GitHub
parent 7afe0faab1
commit 6341d43043
4 changed files with 311 additions and 1 deletions
@@ -26,6 +26,7 @@ from vllm.model_executor.layers.quantization.quark.quark_moe import ( # noqa: E
from vllm.model_executor.layers.quantization.quark.schemes import (
QuarkOCP_MX,
QuarkScheme,
QuarkW4A8_MXFP4_FP8,
QuarkW8A8Fp8,
QuarkW8A8Int8,
)
@@ -350,6 +351,31 @@ class QuarkConfig(QuantizationConfig):
# Only symmetric weight quantization supported.
return is_int8_dtype and is_tensor and is_weight_symmetric and is_static
def _is_w4a8_mxfp4_fp8(
self,
weight_quant: dict[str, Any] | None,
input_quant: dict[str, Any] | None,
) -> bool:
if weight_quant is None or input_quant is None:
return False
is_weight_mxfp4 = (
weight_quant.get("dtype") == "fp4"
and weight_quant.get("qscheme") == "per_group"
and weight_quant.get("group_size") == 32
and weight_quant.get("scale_format") == "e8m0"
and not weight_quant.get("is_dynamic")
)
is_input_fp8 = (
input_quant.get("dtype") == "fp8_e4m3"
and input_quant.get("qscheme") == "per_tensor"
and not input_quant.get("is_dynamic") # Static per-tensor
and input_quant.get("symmetric") is True # Symmetric quantization
)
return is_weight_mxfp4 and is_input_fp8
def _is_w_ocp_mx_a_x(
self, weight_quant: dict[str, Any] | None, input_quant: dict[str, Any] | None
) -> bool:
@@ -504,6 +530,12 @@ class QuarkConfig(QuantizationConfig):
is_static_input_scheme=True,
input_symmetric=input_config.get("symmetric"),
)
elif self._is_w4a8_mxfp4_fp8(weight_config, input_config):
is_w4a8_supported = self._check_scheme_supported(
QuarkW4A8_MXFP4_FP8.get_min_capability(), error=False
)
if is_w4a8_supported:
return QuarkW4A8_MXFP4_FP8(weight_config, input_config)
elif self._is_w_ocp_mx_a_x(weight_config, input_config):
return QuarkOCP_MX(
weight_config, input_config, dynamic_mxfp4_quant=dynamic_mxfp4_quant