Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

Remove RoutedExperts since the bug is fixed. (#2098)

Xin He8b0923c85bfb
diff --git a/auto_round/compressors/model_free.py b/auto_round/compressors/model_free.py
index 8c171a82a..7f9c1f06c 100644
--- a/auto_round/compressors/model_free.py
+++ b/auto_round/compressors/model_free.py
@@ -1094,8 +1094,6 @@ def _build_mxfp_quantization_config(
qconfig = initialize_quantization(scheme=scheme_name, ignore=ignore)
if is_fp_default and scheme_groups:
targets = list(quantized_layers)
- if any(".experts." in name for name in quantized_layers):
- targets.insert(0, "RoutedExperts")
qconfig.config_groups["group_0"].targets = targets
qconfig = qconfig.to_dict()
qconfig["format"] = fmt
@@ -1122,10 +1120,6 @@ def _build_mxfp_quantization_config(
fmt = "mxfp4-pack-quantized" if group_bits == 4 else "mxfp8-quantized"
is_default_group = group_bits == bits
targets = ["Linear"] if is_default_group else layer_names
- # vLLM MoE: prepend RoutedExperts so vLLM's routed-expert matcher
- # takes priority when this explicit group contains expert layers.
- if not is_default_group and any(".experts." in n for n in layer_names):
- targets = ["RoutedExperts"] + targets
tmp_qconfig = initialize_quantization(scheme=scheme_name, ignore=ignore)
group_scheme = tmp_qconfig.config_groups["group_0"]
group_scheme.targets = targets
diff --git a/test/test_cpu/quantization/test_model_free.py b/test/test_cpu/quantization/test_model_free.py
index 8849281b9..520c05d0b 100644
--- a/test/test_cpu/quantization/test_model_free.py
+++ b/test/test_cpu/quantization/test_model_free.py
@@ -640,47 +640,6 @@ class TestModelFreeMXFP:
mxfp4_group = next(g for g in cfg["config_groups"].values() if g["format"] == "mxfp4-pack-quantized")
assert mxfp4_group["targets"] == ["Linear"]
- @require_compressed_tensors
- def test_build_mxfp_mixed_config_adds_routedexperts_for_expert_group(self):
- """Expert layers in a non-Linear group should get RoutedExperts prepended."""
- default = {"bits": 4, "group_size": 32, "sym": True, "data_type": "mx_fp"}
- layer_config = {
- "model.layers.0.mlp.experts.0.down_proj": {"bits": 8, "group_size": 32, "data_type": "mx_fp"},
- "model.layers.0.mlp.experts.1.down_proj": {"bits": 8, "group_size": 32, "data_type": "mx_fp"},
- }
- quantized = [
- "model.layers.0.mlp.experts.0.down_proj",
- "model.layers.0.mlp.experts.1.down_proj",
- "model.layers.0.mlp.gate_proj",
- ]
- cfg = _build_mxfp_quantization_config(default, quantized, ignored_layers=[], layer_config=layer_config)
-
- assert cfg["format"] == "mixed-precision"
- mxfp8_group = next(g for g in cfg["config_groups"].values() if g["format"] == "mxfp8-quantized")
- # RoutedExperts must be first
- assert mxfp8_group["targets"][0] == "RoutedExperts"
- assert "model.layers.0.mlp.experts.0.down_proj" in mxfp8_group["targets"]
- assert "model.layers.0.mlp.experts.1.down_proj" in mxfp8_group["targets"]
- # default MXFP4 group must NOT have RoutedExperts
- mxfp4_group = next(g for g in cfg["config_groups"].values() if g["format"] == "mxfp4-pack-quantized")
- assert "RoutedExperts" not in mxfp4_group["targets"]
-
- @require_compressed_tensors
- def test_build_mxfp_mixed_config_no_routedexperts_without_expert_layers(self):
- """Non-expert explicit group must not get RoutedExperts."""
- default = {"bits": 4, "group_size": 32, "sym": True, "data_type": "mx_fp"}
- layer_config = {
- "model.layers.0.self_attn.q_proj": {"bits": 8, "group_size": 32, "data_type": "mx_fp"},
- }
- quantized = [
- "model.layers.0.self_attn.q_proj",
- "model.layers.0.mlp.fc1",
- ]
- cfg = _build_mxfp_quantization_config(default, quantized, ignored_layers=[], layer_config=layer_config)
-
- mxfp8_group = next(g for g in cfg["config_groups"].values() if g["format"] == "mxfp8-quantized")
- assert "RoutedExperts" not in mxfp8_group["targets"]
-
@require_compressed_tensors
def test_e2e_mxfp_mixed(self, tmp_path):
"""End-to-end: default MXFP4 with some layers overridden to MXFP8."""