From 42c46e6334a9eac123aa00724870deadcd752c4a Mon Sep 17 00:00:00 2001 From: Bi Xue Date: Sun, 29 Mar 2026 08:04:20 -0700 Subject: [PATCH] [sgl] disable piecewise cuda graph when a model doesn't have layers (#21565) --- python/sglang/srt/model_executor/model_runner.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py index fd7b709fb..1a9579f96 100644 --- a/python/sglang/srt/model_executor/model_runner.py +++ b/python/sglang/srt/model_executor/model_runner.py @@ -2383,6 +2383,14 @@ class ModelRunner(ModelRunnerKVCacheMixin): # Collect attention layers and moe layers from the model self.model.model = resolve_language_model(self.model) language_model = getattr(self.model, "language_model", self.model) + + # Some draft models (e.g. eagle3) don't have a standard 'layers' attribute + if not hasattr(language_model.model, "layers"): + logger.warning( + "Disable piecewise CUDA graph because the model does not have a 'layers' attribute" + ) + return + self.attention_layers = [] self.moe_layers = [] self.moe_fusions = []