adapt voxtral (#31095)

Signed-off-by: Patrick von Platen <patrick.v.platen@gmail.com>

vllm/model_executor/models/registry.py

@@ -46,6 +46,7 @@ from .interfaces import (
     has_noops,
     is_attention_free,
     is_hybrid,
+    requires_raw_input_tokens,
     supports_cross_encoding,
     supports_mamba_prefix_caching,
     supports_multimodal,
@@ -422,6 +423,7 @@ _MULTIMODAL_MODELS = {
     ),
     "UltravoxModel": ("ultravox", "UltravoxModel"),
     "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
+    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
     # [Encoder-decoder]
     "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
 }
@@ -539,6 +541,7 @@ class _ModelInfo:
     supports_cross_encoding: bool
     supports_multimodal: bool
     supports_multimodal_raw_input_only: bool
+    requires_raw_input_tokens: bool
     supports_multimodal_encoder_tp_data: bool
     supports_pp: bool
     has_inner_state: bool
@@ -562,6 +565,7 @@ class _ModelInfo:
             supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                 model
             ),
+            requires_raw_input_tokens=requires_raw_input_tokens(model),
             supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                 model
             ),