[Quantization] Remove dead QuantizationConfig.is_mxfp4_quant (#51793)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-08-13 13:58:31 -04:00
committed by GitHub
parent b96bcd0b47
commit 48825acc23
4 changed files with 0 additions and 41 deletions
@@ -258,19 +258,3 @@ class QuantizationConfig(ABC):
# TODO: revision is never passed currently in vllm.py,
# but is used in subclasses, should we remove this parameter?
pass
def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
"""
Determine if mxfp4 quantization will be used for this config.
This allows hidden_size rounding to happen before moe_config creation
without needing to instantiate quant_method first.
Args:
prefix: The layer prefix/name in the model
layer: The layer module
Returns:
True if this config uses MXFP4 quantization, False otherwise
"""
return False
@@ -104,10 +104,6 @@ class Mxfp4Config(QuantizationConfig):
)
return None
def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
"""MXFP4 config always uses MXFP4 quantization."""
return True
class GptOssMxfp4Config(Mxfp4Config):
"""MXFP4 config for GPT-OSS checkpoints.
@@ -576,22 +576,6 @@ class QuarkConfig(QuantizationConfig):
return True
def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
"""
For Quark, determine if it's OCP MXFP4 by checking config directly.
This allows hidden_size rounding to happen before moe_config creation.
"""
layer_quant_config = self._find_matched_config(prefix, layer)
weight_config = layer_quant_config.get("weight")
input_config = layer_quant_config.get("input_tensors")
return (
self._is_w_ocp_mx_a_x(weight_config, input_config)
and weight_config is not None
and weight_config.get("dtype") == "fp4"
and getattr(torch, "float4_e2m1fn_x2", None) is not None
)
def _find_matched_config(
self, layer_name: str, module: torch.nn.Module
) -> dict[str, Any]:
-5
View File
@@ -192,8 +192,3 @@ class DeepseekV4FP8Config(Fp8Config):
# expert_dtype == "fp8": fall through to Fp8Config which
# returns Fp8MoEMethod with block-wise float32 scales.
return super().get_quant_method(layer, prefix)
def is_mxfp4_quant(self, prefix, layer):
if not isinstance(layer, RoutedExperts) or self.expert_dtype != "fp4":
return False
return self.moe_quant_algo != "NVFP4"