[Quantization] Remove dead `QuantizationConfig.is_mxfp4_quant` (#51793)
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
fxmarty-amd committed
Aug 13, 2026 at 19:58 UTC
48825acc232f06380f15124b9a1ea7076939e88c
4 files changed
-41
vllm/model_executor/layers/quantization/base_config.py
-16
@@ -258,19 +258,3 @@ class QuantizationConfig(ABC):
258
# TODO: revision is never passed currently in vllm.py,
259
# but is used in subclasses, should we remove this parameter?
260
pass
261
-
262
- def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
263
- """
264
- Determine if mxfp4 quantization will be used for this config.
265
-
266
- This allows hidden_size rounding to happen before moe_config creation
267
- without needing to instantiate quant_method first.
268
-
269
- Args:
270
- prefix: The layer prefix/name in the model
271
- layer: The layer module
272
-
273
- Returns:
274
- True if this config uses MXFP4 quantization, False otherwise
275
- """
276
- return False
vllm/model_executor/layers/quantization/mxfp4.py
-4
@@ -104,10 +104,6 @@ class Mxfp4Config(QuantizationConfig):
104
)
105
return None
106
107
- def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
108
- """MXFP4 config always uses MXFP4 quantization."""
109
- return True
110
-
107
108
class GptOssMxfp4Config(Mxfp4Config):
109
"""MXFP4 config for GPT-OSS checkpoints.
vllm/model_executor/layers/quantization/quark/quark.py
-16
@@ -576,22 +576,6 @@ class QuarkConfig(QuantizationConfig):
576
577
return True
578
579
- def is_mxfp4_quant(self, prefix: str, layer: torch.nn.Module) -> bool:
580
- """
581
- For Quark, determine if it's OCP MXFP4 by checking config directly.
582
- This allows hidden_size rounding to happen before moe_config creation.
583
- """
584
- layer_quant_config = self._find_matched_config(prefix, layer)
585
- weight_config = layer_quant_config.get("weight")
586
- input_config = layer_quant_config.get("input_tensors")
587
-
588
- return (
589
- self._is_w_ocp_mx_a_x(weight_config, input_config)
590
- and weight_config is not None
591
- and weight_config.get("dtype") == "fp4"
592
- and getattr(torch, "float4_e2m1fn_x2", None) is not None
593
- )
594
-
579
def _find_matched_config(
580
self, layer_name: str, module: torch.nn.Module
581
) -> dict[str, Any]:
vllm/models/deepseek_v4/quant_config.py
-5
@@ -192,8 +192,3 @@ class DeepseekV4FP8Config(Fp8Config):
192
# expert_dtype == "fp8": fall through to Fp8Config which
193
# returns Fp8MoEMethod with block-wise float32 scales.
194
return super().get_quant_method(layer, prefix)
195
-
196
- def is_mxfp4_quant(self, prefix, layer):
197
- if not isinstance(layer, RoutedExperts) or self.expert_dtype != "fp4":
198
- return False
199
- return self.moe_quant_algo != "NVFP4"