From e60f3c4c137f23bb870524b937719d24bde8715b Mon Sep 17 00:00:00 2001 From: Tan Pin Siang Date: Wed, 12 Aug 2026 03:52:30 -0700 Subject: [PATCH] [Bugfix] Fix MiniMax-M3 compressed-tensors FP8 MoE SwiGLU params (#46845) Signed-off-by: Tan Pin Siang --- tests/quantization/test_compressed_tensors.py | 32 +++++++++++++++++++ .../compressed_tensors_moe_w8a8_fp8.py | 2 ++ 2 files changed, 34 insertions(+) diff --git a/tests/quantization/test_compressed_tensors.py b/tests/quantization/test_compressed_tensors.py index 19f63f2286..3f11cec1c1 100644 --- a/tests/quantization/test_compressed_tensors.py +++ b/tests/quantization/test_compressed_tensors.py @@ -22,6 +22,7 @@ from vllm.model_executor.kernels.linear import ( Fp8BlockScaledMMLinearKernel, ) from vllm.model_executor.layers.fused_moe import UnquantizedFusedMoEMethod +from vllm.model_executor.layers.fused_moe.oracle.fp8 import Fp8MoeBackend from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501 CompressedTensorsConfig, CompressedTensorsLinearMethod, @@ -35,6 +36,9 @@ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tenso CompressedTensorsWNA8O8Int, CompressedTensorsWNA16, ) +from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe.compressed_tensors_moe_w8a8_fp8 import ( # noqa: E501 + CompressedTensorsW8A8Fp8MoEMethod, +) from vllm.model_executor.layers.quantization.compressed_tensors.utils import ( find_matched_target, ) @@ -345,6 +349,34 @@ def test_compressed_tensors_fp8(vllm_runner): assert output +def test_compressed_tensors_w8a8_fp8_moe_forwards_swiglu_params(): + quant_method = object.__new__(CompressedTensorsW8A8Fp8MoEMethod) + quant_method.input_quant = QuantizationArgs( + num_bits=8, + type=QuantizationType.FLOAT, + strategy=QuantizationStrategy.TOKEN, + dynamic=True, + symmetric=True, + ) + quant_method.weight_block_size = None + quant_method.fp8_backend = Fp8MoeBackend.TRITON + + layer = Mock() + layer.w13_weight_scale = torch.ones(2, 4, 1) + layer.w2_weight_scale = torch.ones(2, 8, 1) + layer.w13_input_scale = None + layer.w2_input_scale = None + layer.swiglu_alpha = 1.702 + layer.swiglu_beta = None + layer.swiglu_limit = 7.0 + + quant_config = quant_method.get_fused_moe_quant_config(layer) + + assert quant_config.gemm1_alpha == 1.702 + assert quant_config.gemm1_beta is None + assert quant_config.gemm1_clamp_limit == 7.0 + + @pytest.mark.skipif( not current_platform.is_cuda(), reason="This test is skipped on non-CUDA platform." ) diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w8a8_fp8.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w8a8_fp8.py index 704bc9a87c..7829ef65c0 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w8a8_fp8.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w8a8_fp8.py @@ -352,6 +352,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod): per_act_token_quant=is_per_token, per_out_ch_quant=is_per_token, block_shape=self.weight_block_size, + gemm1_alpha=getattr(layer, "swiglu_alpha", None), + gemm1_beta=getattr(layer, "swiglu_beta", None), swiglu_limit=getattr(layer, "swiglu_limit", None), layer=layer, )