[Cookbook] Nemotron3-Ultra: Add mamba-backend and SSM dtype flags (#28675)

Co-authored-by: Brayden Zhong <brayden.zhong@radixark.ai>

docs_new/cookbook/autoregressive/NVIDIA/Nemotron3-Ultra.mdx

@@ -60,12 +60,20 @@ The generator only emits a runnable command for combinations that NVIDIA / SGLan
 - **Attention backend**:
 
     **H100/H200**: Use flash attention 3 backend by default.
-    **B200/GB200/B300/GB300**: Append `--attention-backend trtllm_mha`. The flashinfer default breaks the overlap scheduler on Blackwell, so `trtllm_mha` is required there.
+    **B200/GB200/B300/GB300**: Set `--attention-backend trtllm_mha`. The flashinfer default breaks the overlap scheduler on Blackwell, so `trtllm_mha` is required there.
 
 - **Mamba scheduler strategy**:
 
     Always launch with `--mamba-scheduler-strategy extra_buffer`. This hybrid Transformer-Mamba model requires the `extra_buffer` strategy for correct scheduling of its Mamba state.
 
+- **Mamba backend**:
+
+    The Mamba layers use the Triton SSM kernels by default. For better performance, set `--mamba-backend flashinfer` to use the FlashInfer Mamba kernels instead.
+
+- **Mamba SSM precision**:
+
+    The SSM state dtype defaults to the model config value. Set `--mamba-ssm-dtype float16` to store the Mamba states in FP16, which reduces mamba cache memory without significant accuracy loss.
+
 - **TP support**:
 
     To set tp size, use `--tp <4|8|16>`. Recommended pairings:
@@ -82,7 +90,7 @@ The generator only emits a runnable command for combinations that NVIDIA / SGLan
 
 - **Expert parallel (EP)**:
 
-    This MoE only supports `ep_size == 1` (off) or `ep_size == tp_size`. To enable expert parallelism, append `--ep <tp>` with the same value as `--tp`.
+    This MoE only supports `ep_size == 1` (off) or `ep_size == tp_size`. To enable expert parallelism, set `--ep <tp>` with the same value as `--tp`.
 
 - **Multi-token prediction (MTP)**:
 
@@ -90,15 +98,15 @@ The generator only emits a runnable command for combinations that NVIDIA / SGLan
 
 - **FP8 KV cache**:
 
-    To enable fp8 kv cache, please append `--kv-cache-dtype fp8_e4m3`.
+    To enable fp8 kv cache, set `--kv-cache-dtype fp8_e4m3`. This is enabled by default on the NVFP4 checkpoint.
 
 - **Reasoning parser**:
 
-    Append `--reasoning-parser nemotron_3` to enable structured reasoning traces (`reasoning_content` field in the response).
+    Set `--reasoning-parser nemotron_3` to enable structured reasoning traces (`reasoning_content` field in the response).
 
 - **Tool calling**:
 
-    Append `--tool-call-parser qwen3_coder` to enable tool calling support.
+    Set `--tool-call-parser qwen3_coder` to enable tool calling support.
 
 ## 4. Model Invocation
 

docs_new/src/snippets/autoregressive/nemotron3-ultra-deployment.jsx

@@ -154,6 +154,24 @@ export const Nemotron3UltraDeployment = () => {
         { id: 'bf16',     label: 'bf16',     default: false }
       ]
     },
+    mambabackend: {
+      name: 'mambabackend',
+      title: 'Mamba Backend',
+      items: [
+        { id: 'triton',     label: 'Triton',     subtitle: 'Default', default: true  },
+        { id: 'flashinfer', label: 'FlashInfer', subtitle: 'Faster',  default: false }
+      ],
+      commandRule: (value) => value === 'flashinfer' ? '--mamba-backend flashinfer' : null
+    },
+    mambassmdtype: {
+      name: 'mambassmdtype',
+      title: 'Mamba SSM DType',
+      items: [
+        { id: 'default', label: 'Default', subtitle: 'Model config', default: true  },
+        { id: 'float16', label: 'float16', subtitle: 'Less memory',   default: false }
+      ],
+      commandRule: (value) => value === 'float16' ? '--mamba-ssm-dtype float16' : null
+    },
     thinking: {
       name: 'thinking',
       title: 'Reasoning Parser',