From 8fbcfd07236df523fe2bc254ba394b8ca49b5bf5 Mon Sep 17 00:00:00 2001 From: Ke Bao Date: Fri, 1 Aug 2025 00:49:26 +0800 Subject: [PATCH] Update step3v default config (#8626) --- python/sglang/srt/configs/model_config.py | 1 + python/sglang/srt/models/step3_vl.py | 3 --- python/sglang/srt/utils.py | 1 + 3 files changed, 2 insertions(+), 3 deletions(-) diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py index 37fbf07c7..f3643d154 100644 --- a/python/sglang/srt/configs/model_config.py +++ b/python/sglang/srt/configs/model_config.py @@ -112,6 +112,7 @@ class ModelConfig: mm_disabled_models = [ "Gemma3ForConditionalGeneration", "Llama4ForConditionalGeneration", + "Step3VLForConditionalGeneration", ] if self.hf_config.architectures[0] in mm_disabled_models: enable_multimodal = False diff --git a/python/sglang/srt/models/step3_vl.py b/python/sglang/srt/models/step3_vl.py index 3ed0a153f..c43d5bec8 100644 --- a/python/sglang/srt/models/step3_vl.py +++ b/python/sglang/srt/models/step3_vl.py @@ -868,7 +868,6 @@ class Step3VLForConditionalGeneration(nn.Module): ) def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): - # TODO: stacked_params_mapping = [ # (param_name, shard_name, shard_id) (".qkv_proj", ".q_proj", 0), @@ -901,9 +900,7 @@ class Step3VLForConditionalGeneration(nn.Module): for name, loaded_weight in weights: if "vision_model" in name: - # 1.It’s not great, but let’s leave it like this for now name = name.replace("self_attn", "self_attn.attn") - # 2. name = name.replace("out_proj", "proj") # TODO: support vision model diff --git a/python/sglang/srt/utils.py b/python/sglang/srt/utils.py index f824a006a..33112052b 100644 --- a/python/sglang/srt/utils.py +++ b/python/sglang/srt/utils.py @@ -2344,6 +2344,7 @@ def is_fa3_default_architecture(hf_config): "Qwen3ForCausalLM", "Qwen3MoeForCausalLM", "Glm4MoeForCausalLM", + "Step3VLForConditionalGeneration", } return architectures[0] in default_archs