From fe9d85d93cbd22283809317c61df6a9f5df35f90 Mon Sep 17 00:00:00 2001 From: Alison Shao <54658187+alisonshao@users.noreply.github.com> Date: Mon, 2 Mar 2026 19:03:21 -0800 Subject: [PATCH] Fix CompressedTensorsMxInt4MoE abstract method and relax GPQA baseline (#19726) Co-authored-by: Alison Shao --- .../schemes/compressed_tensors_w4a4_mxint4_moe.py | 2 +- test/registered/8-gpu-models/test_deepseek_v32.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_mxint4_moe.py b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_mxint4_moe.py index 8419efcb8..865f3de43 100644 --- a/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_mxint4_moe.py +++ b/python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_mxint4_moe.py @@ -287,7 +287,7 @@ class CompressedTensorsMxInt4MoE(CompressedTensorsMoEScheme): ): self.moe_runner_config = moe_runner_config - def apply( + def apply_weights( self, layer: torch.nn.Module, dispatch_output: StandardDispatchOutput, diff --git a/test/registered/8-gpu-models/test_deepseek_v32.py b/test/registered/8-gpu-models/test_deepseek_v32.py index 154fea493..5a6525d20 100644 --- a/test/registered/8-gpu-models/test_deepseek_v32.py +++ b/test/registered/8-gpu-models/test_deepseek_v32.py @@ -30,7 +30,7 @@ DP_ARGS = [ # Accuracy thresholds GSM8K_BASELINE = 0.935 -GPQA_BASELINE = 0.835 +GPQA_BASELINE = 0.83 class TestDeepseekV32(unittest.TestCase):