[None][chore] Update AD model list (#14686)

Signed-off-by: Yueh-Ting (eop) Chen <yuehtingc@nvidia.com>
Signed-off-by: tcherckez-nvidia <127761168+tcherckez-nvidia@users.noreply.github.com>
Co-authored-by: Yueh-Ting (eop) Chen <yuehtingc@nvidia.com>

Author: tcherckez-nvidia

examples/auto_deploy/model_registry/configs/granite_4.0_h_small.yaml

@@ -7,7 +7,8 @@
 # MoE experts distributed via EP across 4 GPUs. This avoids the TP incompatibility
 # where n_groups=1 makes Mamba B/C SSM dimensions non-shardable.
 attn_backend: flashinfer
-max_batch_size: 16
+cuda_graph_config:
+  max_batch_size: 16
 kv_cache_config:
   free_gpu_memory_fraction: 0.5
 transforms:

examples/auto_deploy/model_registry/configs/granite_4.0_micro.yaml

@@ -4,6 +4,7 @@
 # Configuration for granite-4.0-micro (and similar GraniteMoeHybrid models)
 # Uses flashinfer attention backend (trtllm backend produces garbled output).
 attn_backend: flashinfer
-max_batch_size: 16
+cuda_graph_config:
+  max_batch_size: 16
 kv_cache_config:
   free_gpu_memory_fraction: 0.5

examples/auto_deploy/model_registry/configs/granite_4.0_tiny_preview.yaml

@@ -5,7 +5,8 @@
 # Uses flashinfer attention backend (trtllm backend produces garbled output with muP scaling)
 # dtype is missing from config for this model, so we provide it explicitly.
 attn_backend: flashinfer
-max_batch_size: 16
+cuda_graph_config:
+  max_batch_size: 16
 kv_cache_config:
   free_gpu_memory_fraction: 0.5
 model_kwargs:

examples/auto_deploy/model_registry/configs/hunyuan_a13b_ep.yaml

@@ -2,7 +2,8 @@
 # SPDX-License-Identifier: Apache-2.0
 
 # Hunyuan A13B MoE: replicate attention/dense layers and shard routed experts with EP.
-max_batch_size: 16
+cuda_graph_config:
+  max_batch_size: 16
 kv_cache_config:
   free_gpu_memory_fraction: 0.5
 transforms:

examples/auto_deploy/model_registry/configs/qwen3Next.yaml

@@ -7,6 +7,8 @@ attn_backend: torch
 max_seq_len: 4096
 max_num_tokens: 4096
 max_batch_size: 64
+cuda_graph_config:
+  max_batch_size: 64
 enable_chunked_prefill: true
 model_factory: AutoModelForCausalLM
 # disable_overlap_scheduler: false