[feat][sgl-atom] add qwen3 sglang dense model support (#1416)

* fix Qwen3 dense startup with ATOM backend

Register Qwen3 dense for the SGLang ATOM wrapper and route page-size 1 dense decode through the existing native AITER path to avoid invalid KV layout reshapes during CUDA graph capture.

Co-authored-by: Cursor <cursoragent@cursor.com>

* Limit Qwen3 dense startup fix scope

Restore the existing dense attention routing and use page-size 16 only for Qwen3-32B MI308 SGLang benchmark entries so the model stays on the ATOM attention path.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Hexiang Wang <56632993+whx-sjtu@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: 3 people

.github/benchmark/sglang_benchmark_models.json

@@ -955,6 +955,7 @@
       "prefix": "qwen3-32b-fp8-tp1-mi308",
       "extra_args": [
         "--tensor-parallel-size 1",
+        "--page-size 16",
         "qwen_reasoning"
       ],
       "bench_args": "",
@@ -971,6 +972,7 @@
       "prefix": "qwen3-32b-fp8-tp8-mi308",
       "extra_args": [
         "--tensor-parallel-size 8",
+        "--page-size 16",
         "qwen_reasoning"
       ],
       "bench_args": "",

atom/plugin/sglang/runtime/model_arch.py

@@ -86,6 +86,7 @@ def _install_deepseek_v4_adapters(model: Any) -> None:
         prepare_config=_prepare_kimi_k25_config,
         install_adapters=_install_deepseek_mla_adapters,
     ),
+    "Qwen3ForCausalLM": SGLangModelAdapterSpec(),
     "Qwen3MoeForCausalLM": SGLangModelAdapterSpec(),
     "Qwen3NextForCausalLM": SGLangModelAdapterSpec(
         wrapper_binds_gdn_context=True,
@@ -114,6 +115,7 @@ def _install_deepseek_v4_adapters(model: Any) -> None:
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
         "GlmMoeDsaForCausalLM",
+        "Qwen3ForCausalLM",
         "Qwen3MoeForCausalLM",
         "Qwen3NextForCausalLM",
         "MiniMaxM2ForCausalLM",