PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
enable MXFP clamp for model free (#1914)
Xin He50ee58c9e176
diff --git a/auto_round/compressors/model_free.py b/auto_round/compressors/model_free.pyindex 8717ab17c..362ca7b51 100644--- a/auto_round/compressors/model_free.py+++ b/auto_round/compressors/model_free.py@@ -520,7 +520,7 @@ def _quantize_weight_mxfp(# quant_mx returns (qdq_tensor, shared_exp, None). We only need shared_exp# (the per-block log2 scale). The element-wise rounding to the FP4/FP8 grid# is performed inside QuantLinear.pack via dtype casts / pack_fp4_to_uint8.- _, shared_exp, _ = quant_mx(weight_dev, bits=bits, group_size=group_size, data_type=data_type)+ weight_dev, shared_exp, _ = quant_mx(weight_dev, bits=bits, group_size=group_size, data_type=data_type)# Reshape to (out_features, n_groups) so the on-disk weight_scale matches# the llm-compressor convention (and QuantLinear's registered buffer shape).shared_exp = shared_exp.reshape(out_features, in_features // group_size)diff --git a/test/test_cpu/quantization/test_model_free.py b/test/test_cpu/quantization/test_model_free.pyindex daa82c617..d8e45336c 100644--- a/test/test_cpu/quantization/test_model_free.py+++ b/test/test_cpu/quantization/test_model_free.py@@ -351,6 +351,23 @@ class TestModelFreeMXFP:assert out["layer.weight_scale"].shape == (64, 4)assert out["layer.weight_scale"].dtype == torch.uint8+ def test_mxfp8_max_value_scaling(self):+ """Input bfloat16 weight with max value 240; after MXFP8 quantization+ the stored FP8 values should be clamped to 448 (fp8_e4m3fn max) and+ must NOT be NaN.++ quant_mx computes shared_exp = floor(log2(240)) - emax = 7 - 8 = -1,+ so scale = 2^-1 = 0.5. The normalised value is 240/0.5 = 480 which+ exceeds the FP8 max of 448 and must be clamped before the fp8 cast.+ """+ w_bf16 = torch.full((64, 128), 240, dtype=torch.bfloat16)+ out = _quantize_weight_mxfp(w_bf16, "layer", bits=8, group_size=32, data_type="mx_fp")+ assert out["layer.weight"].dtype == torch.float8_e4m3fn+ weight_fp32 = out["layer.weight"].to(torch.float32)+ assert not torch.isnan(weight_fp32).any(), "FP8 weight must not contain NaN"+ max_val = torch.max(weight_fp32).item()+ assert max_val == 448, f"Expected 448 (fp8_e4m3fn max after clamp), got {max_val}"+@require_compressed_tensors@pytest.mark.parametrize("scheme,fmt", [("MXFP4", "mxfp4-pack-quantized"), ("MXFP8", "mxfp8-quantized")])def test_e2e_mxfp(self, tmp_path, scheme, fmt):