support qwen3_next blackwell (#10403)

python/sglang/srt/layers/attention/triton_backend.py

@@ -80,7 +80,13 @@ class TritonAttnBackend(AttentionBackend):
         self.num_kv_head = model_runner.model_config.get_num_kv_heads(
             get_attention_tp_size()
         )
-        self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[-1]
+        if model_runner.is_hybrid_gdn:
+            # For hybrid linear models, layer_id = 0 may not be full attention
+            self.v_head_dim = model_runner.token_to_kv_pool.get_v_head_dim()
+        else:
+            self.v_head_dim = model_runner.token_to_kv_pool.get_value_buffer(0).shape[
+                -1
+            ]
         self.max_context_len = model_runner.model_config.context_len
         self.device = model_runner.device
         self.device_core_count = get_device_core_count(model_runner.gpu_id)

python/sglang/srt/mem_cache/memory_pool.py

@@ -728,6 +728,9 @@ class HybridLinearKVPool(KVCache):
             layer_id_override=layer_id,
         )
 
+    def get_v_head_dim(self):
+        return self.full_kv_pool.get_value_buffer(0).shape[-1]
+
 
 class SWAKVPool(KVCache):
     """KV cache with separate pools for full and SWA attention layers."""

python/sglang/srt/model_executor/model_runner.py

@@ -127,6 +127,7 @@ from sglang.srt.utils import (
     get_bool_env_var,
     get_cpu_ids_by_node,
     init_custom_process_group,
+    is_blackwell,
     is_fa3_default_architecture,
     is_flashinfer_available,
     is_hip,
@@ -1832,6 +1833,10 @@ class ModelRunner:
                 from sglang.srt.layers.attention.ascend_backend import AscendAttnBackend
 
                 full_attn_backend = AscendAttnBackend(self)
+            elif is_blackwell():
+                from sglang.srt.layers.attention.triton_backend import TritonAttnBackend
+
+                full_attn_backend = TritonAttnBackend(self)
             else:
                 from sglang.srt.layers.attention.flashattention_backend import (
                     FlashAttentionBackend,

python/sglang/srt/speculative/eagle_worker.py

@@ -48,6 +48,7 @@ from sglang.srt.utils import (
     empty_context,
     get_available_gpu_memory,
     get_bool_env_var,
+    is_blackwell,
     is_cuda,
     next_power_of_2,
 )
@@ -214,7 +215,11 @@ class EAGLEWorker(TpModelWorker):
             "triton": self._create_triton_decode_backend,
             "aiter": self._create_aiter_decode_backend,
             "fa3": self._create_fa3_decode_backend,
-            "hybrid_linear_attn": self._create_fa3_decode_backend,
+            "hybrid_linear_attn": (
+                self._create_fa3_decode_backend
+                if not is_blackwell()
+                else self._create_triton_decode_backend
+            ),
             "flashmla": self._create_flashmla_decode_backend,
             "trtllm_mha": self._create_trtllm_mha_decode_backend,
             "trtllm_mla": self._create_trtllm_mla_decode_backend,
@@ -232,7 +237,11 @@ class EAGLEWorker(TpModelWorker):
             "triton": self._create_triton_prefill_backend,
             "aiter": self._create_aiter_prefill_backend,
             "fa3": self._create_fa3_prefill_backend,
-            "hybrid_linear_attn": self._create_fa3_prefill_backend,
+            "hybrid_linear_attn": (
+                self._create_fa3_prefill_backend
+                if not is_blackwell()
+                else self._create_triton_prefill_backend
+            ),
             "trtllm_mha": self._create_trtllm_mha_prefill_backend,
             "trtllm_mla": self._create_trtllm_mla_prefill_backend,
         }