[AMD] Prefer HIP Top-K for GLM-5.x on ROCm (#39631)

python/sglang/srt/arg_groups/model_hook.py

@@ -340,24 +340,9 @@ def handle_model_specific_adjustments(server_args: Any):
 
         run_post_process_pass(server_args, _deepseek_moe_quant_resolution)
         if get_platform().is_hip:
-            if is_deepseek_dsa(hf_config):
-                # The fused top-k v2 kernel (topk_transform_paged_v2) is a
-                # CUDA/Hopper-only path: its JIT source includes
-                # <cooperative_groups.h> and uses cg::this_cluster()
-                # (thread-block clusters), neither of which exists on ROCm,
-                # so it fails to JIT-compile on gfx9xx during CUDA-graph
-                # capture. DeepSeek-V4 already disables it on HIP; mirror that
-                # here for the rest of the DSA family (DeepSeek-V3.2 /
-                # GLM-5.x) that shares the same decode top-k path.
+            if is_deepseek_dsa(hf_config) and not envs.SGLANG_OPT_USE_TOPK_V2.is_set():
+                # Prefer HIP top-k by default while honoring an explicit selection.
                 envs.SGLANG_OPT_USE_TOPK_V2.set(False)
-            if model_arch == "GlmMoeDsaForCausalLM":
-                # Open the fused top-k v2 kernel for the GLM-5.x DSA
-                # family on ROCm: it shares this decode top-k path, and
-                # the kernel's ROCm build compiles the streaming levels
-                # on gfx9xx. Order is load-bearing: the blanket disable
-                # above `set`s the variable unconditionally, so this has
-                # to follow it.
-                envs.SGLANG_OPT_USE_TOPK_V2.set(True)
             if not resolved_view(server_args).enable_dp_attention and cfg.nnodes == 1:
                 # TODO (Hubert): Put this back later
                 # server_args.enable_aiter_allreduce_fusion = True