[Model] Update Kimi-K25 and Isaac processors to fit HF-style (#37693)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-03-20 18:30:22 +00:00
committed by GitHub
parent d7d2b5e405
commit 37aadf6237
5 changed files with 128 additions and 95 deletions
+3 -4
View File
@@ -334,15 +334,14 @@ class IsaacProcessingInfo(BaseProcessingInfo):
return IsaacConfig()
def get_image_processor(self, **kwargs) -> IsaacImageProcessor:
return IsaacImageProcessor(kwargs)
return IsaacImageProcessor(**kwargs)
def get_hf_processor(self, **kwargs) -> IsaacProcessor:
hf_config = self.get_hf_config()
return self.ctx.init_processor(
IsaacProcessor,
return IsaacProcessor(
tokenizer=self.get_tokenizer(),
image_processor=self.get_image_processor(),
image_processor=self.get_image_processor(**kwargs),
image_token=hf_config.vision_token,
)
+16 -18
View File
@@ -104,19 +104,25 @@ class KimiK25ProcessingInfo(BaseProcessingInfo):
def __init__(self, ctx: InputProcessingContext) -> None:
super().__init__(ctx)
self.hf_config = self.get_hf_config()
self.media_token_id = self.hf_config.media_placeholder_token_id
media_processor = cached_get_image_processor(
self.hf_config = hf_config = self.get_hf_config()
tokenizer = self.get_tokenizer()
image_processor = cached_get_image_processor(
self.ctx.model_config.model,
trust_remote_code=self.ctx.model_config.trust_remote_code,
)
self.media_processor = media_processor
self.media_token_id = media_token_id = hf_config.media_placeholder_token_id
self.media_token = tokenizer.decode(media_token_id)
self.image_processor = image_processor
self.hf_processor = KimiK25Processor(
media_processor=self.media_processor,
tokenizer=self.get_tokenizer(),
media_token_id=self.media_token_id,
tokenizer=tokenizer,
image_processor=image_processor,
media_token_id=media_token_id,
)
self.media_tokens_calculator = self.media_processor.media_tokens_calculator
self.media_tokens_calculator = image_processor.media_tokens_calculator
def get_hf_processor(self):
return self.hf_processor
@@ -132,20 +138,15 @@ class KimiK25ProcessingInfo(BaseProcessingInfo):
class KimiK25DummyInputsBuilder(BaseDummyInputsBuilder[KimiK25ProcessingInfo]):
"""Builds dummy inputs for Kimi-K2.5 model profiling."""
def __init__(self, info: KimiK25ProcessingInfo) -> None:
super().__init__(info)
self.media_token_id = self.info.media_token_id
self.frame_per_chunk = self.info.media_processor.num_frames_per_chunk
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
num_media = mm_counts.get("vision_chunk", 0)
return "<|media_pad|>" * num_media
return self.info.media_token * num_media
def get_dummy_mm_items(self):
dummy_videos = self._get_dummy_images(
height=MaxImageTokenMeta.height,
width=MaxImageTokenMeta.width,
num_images=self.frame_per_chunk,
num_images=self.info.image_processor.num_frames_per_chunk,
)
video_chunk_dummy_item = VisionChunkVideo(
@@ -236,9 +237,6 @@ class KimiK25MultiModalProcessor(BaseMultiModalProcessor[KimiK25ProcessingInfo])
),
]
def split_video_chunks(self, video):
return self.info.media_processor.split_video_chunks(video)
@MULTIMODAL_REGISTRY.register_processor(
KimiK25MultiModalProcessor,