[Fix] Read the granite sinks dtype from the exec bag, not the legacy global shim (#35921)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

python/sglang/srt/models/granite.py

@@ -47,8 +47,7 @@ from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
     sharded_weight_loader,
 )
-from sglang.srt.runtime_context import get_parallel
-from sglang.srt.server_args import get_global_server_args
+from sglang.srt.runtime_context import get_exec, get_parallel
 from sglang.srt.utils import add_prefix, set_weight_attrs
 from sglang.utils import get_exception_traceback
 
@@ -74,8 +73,10 @@ def granite_layer_attn_params(
 
 
 def build_attention_sinks(num_heads: int) -> nn.Parameter:
-    # trtllm_mha requires float32 sinks, other backends use bfloat16.
-    attn_backend = get_global_server_args().attention_backend
+    # TODO(kpham-sgl): one parameter cannot serve a split launch -- trtllm_mha
+    # wants float32 sinks, FA4 wants bfloat16. Pick the dtype at init instead,
+    # once the serving backends are known. Checkpoint dtype also unverified.
+    attn_backend = get_exec().kernel.attention_backend
     sinks_dtype = torch.float32 if attn_backend == "trtllm_mha" else torch.bfloat16
     sinks = nn.Parameter(torch.empty(num_heads, dtype=sinks_dtype), requires_grad=False)
     set_weight_attrs(sinks, {"weight_loader": sharded_weight_loader(0)})