fix: Fix nanov2 nightlies for new vllm (#2636)

Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>

Author: yfw

examples/configs/recipes/llm/grpo-nano-v2-12b-1n8g-megatron.yaml

@@ -21,6 +21,10 @@ policy:
       max_model_len: 512
     vllm_kwargs:
       mamba_ssm_cache_dtype: float32  # needed to avoid SSM state overflows in vllm 5.2.0
+      # vLLM 0.20 allocates one Mamba cache block per decode seq; the default
+      # max_num_seqs=1024 exceeds available blocks. Cap to the rollout batch
+      # (num_prompts_per_step * num_generations_per_prompt = 32 * 16 = 512).
+      max_num_seqs: 512
   sequence_packing:
     enabled: false
 data:

examples/configs/recipes/llm/grpo-nano-v2-12b-2n8g-fsdp2tp1.yaml

@@ -25,6 +25,10 @@ policy:
       max_model_len: 512
     vllm_kwargs:
       mamba_ssm_cache_dtype: float32  # needed to avoid SSM state overflows in vllm 5.2.0
+      # vLLM 0.20 allocates one Mamba cache block per decode seq; the default
+      # max_num_seqs=1024 exceeds available blocks. Cap to the rollout batch
+      # (num_prompts_per_step * num_generations_per_prompt = 32 * 16 = 512).
+      max_num_seqs: 512
   scheduler:
     - name: "torch.optim.lr_scheduler.LinearLR"
       kwargs: