[Recipe] Add Fish Speech S2 Pro 2-GPU deploy profile (#3323)

Signed-off-by: Yueqian Lin <linyueqian@outlook.com>

vllm_omni/deploy/fish_qwen3_omni_2gpu.yaml

@@ -0,0 +1,72 @@
+# Fish Speech S2 Pro 2-GPU deploy: Stage0 AR on GPU 0, Stage1 DAC on GPU 1.
+#
+# Opt-in profile validated on H100x2. Compared with the single-GPU default
+# (``fish_qwen3_omni.yaml``), it splits AR and DAC across devices to remove
+# AR/DAC contention, enlarges Stage0 AR batching, and lets Stage1 consume
+# multiple DAC chunks in parallel.
+#
+# Select via:
+#   CUDA_VISIBLE_DEVICES=0,1 vllm serve fishaudio/s2-pro --omni \
+#       --deploy-config vllm_omni/deploy/fish_qwen3_omni_2gpu.yaml
+#
+# Or programmatically:
+#   Omni(model="fishaudio/s2-pro",
+#        deploy_config_path="vllm_omni/deploy/fish_qwen3_omni_2gpu.yaml")
+
+async_chunk: true
+enable_chunked_prefill: false
+
+connectors:
+  connector_of_shared_memory:
+    name: SharedMemoryConnector
+    extra:
+      shm_threshold_bytes: 65536
+      codec_streaming: true
+      # Tightened from 10ms to 1ms: chunks become available sooner under
+      # cross-GPU pipelining, so connector idle waiting dominates if left high.
+      connector_get_sleep_s: 0.001
+      connector_get_max_wait_first_chunk: 3000
+      connector_get_max_wait: 300
+      # ~21 Hz codec; 25 frames ≈ 1.16s of audio.
+      codec_chunk_frames: 25
+      codec_left_context_frames: 25
+      initial_codec_chunk_frames: 4
+
+stages:
+  - stage_id: 0
+    max_num_seqs: 64
+    gpu_memory_utilization: 0.75
+    enforce_eager: false
+    async_scheduling: false
+    # vLLM scheduler requires max_num_batched_tokens >= max_model_len when
+    # chunked prefill is off (SlowAR decode loop isn't chunked-prefill-safe).
+    max_num_batched_tokens: 16384
+    max_model_len: 16384
+    devices: "0"
+    output_connectors:
+      to_stage_1: connector_of_shared_memory
+    default_sampling_params:
+      temperature: 0.8
+      top_k: 30
+      top_p: 0.9
+      max_tokens: 2048
+      seed: 42
+      repetition_penalty: 1.0
+
+  - stage_id: 1
+    max_num_seqs: 64
+    gpu_memory_utilization: 0.2
+    enforce_eager: true
+    async_scheduling: false
+    max_num_batched_tokens: 16384
+    max_model_len: 16384
+    devices: "1"
+    input_connectors:
+      from_stage_0: connector_of_shared_memory
+    default_sampling_params:
+      temperature: 0.0
+      top_p: 1.0
+      top_k: -1
+      max_tokens: 65536
+      seed: 42
+      repetition_penalty: 1.0