[Feature] timestamps for TTS via a forced-aligner pooling (#4795)

Signed-off-by: wjinxu <1299461899@qq.com>
Co-authored-by: xurui <xurui1@modelbest.onaliyun.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

vllm_omni/entrypoints/cli/serve.py

@@ -237,9 +237,8 @@ class OmniServeCommand(CLISubcommand):
             "VoiceDesign, or Base, while diffusion models may use it to select "
             "task-specific weights. If omitted, the model default is used.",
         )
-        # Forced aligner / word timestamps. --forced-aligner is the opt-in
-        # toggle; heavier knobs (gpu_memory_utilization, dtype, max_model_len)
-        # live in the deploy YAML passed via --forced-aligner-config.
+        # Forced aligner / word timestamps. Either flag opts in (--forced-aligner-config
+        # alone works when its YAML sets the model); heavier knobs live in that YAML.
         omni_config_group.add_argument(
             "--forced-aligner",
             type=str,
@@ -260,6 +259,13 @@ class OmniServeCommand(CLISubcommand):
                 "flag, when set, overrides the YAML model field."
             ),
         )
+        omni_config_group.add_argument(
+            "--forced-aligner-device",
+            type=str,
+            default=None,
+            help="Device(s) for the forced-aligner stage (e.g. '2'). Defaults to "
+            "sharing an existing stage's GPU when unset.",
+        )
         omni_config_group.add_argument(
             "--deploy-config",
             type=str,