[MM] Move get_dummy_processor_inputs into MM processor (#57967)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
@@ -189,38 +189,6 @@ class HCXVisionV2DummyInputsBuilder(BaseDummyInputsBuilder[HCXVisionV2Processing
|
||||
num_videos = mm_counts.get("video", 0)
|
||||
return V2_IMAGE_TOKEN * num_images + V2_VIDEO_TOKEN * num_videos
|
||||
|
||||
def get_dummy_processor_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions | None = None,
|
||||
mm_processor_kwargs: Mapping[str, object] | None = None,
|
||||
) -> ProcessorInputs:
|
||||
"""Build dummy processor inputs for memory profiling."""
|
||||
num_images = mm_counts.get("image", 0)
|
||||
num_videos = mm_counts.get("video", 0)
|
||||
prompt_text = V2_IMAGE_TOKEN * num_images + V2_VIDEO_TOKEN * num_videos
|
||||
|
||||
dummy_mm_data = self.get_dummy_mm_data(
|
||||
seq_len,
|
||||
mm_counts,
|
||||
mm_options,
|
||||
mm_processor_kwargs=mm_processor_kwargs,
|
||||
)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data, validate=False)
|
||||
|
||||
tokenizer = self.info.get_tokenizer()
|
||||
prompt = tokenizer.encode(
|
||||
prompt_text,
|
||||
**self.info.default_tok_params.get_encode_kwargs(),
|
||||
)
|
||||
|
||||
return ProcessorInputs(
|
||||
prompt=prompt,
|
||||
mm_data_items=dummy_mm_items,
|
||||
hf_processor_mm_kwargs=mm_processor_kwargs or {},
|
||||
)
|
||||
|
||||
def get_dummy_mm_data(
|
||||
self,
|
||||
seq_len: int,
|
||||
@@ -255,6 +223,28 @@ class HCXVisionV2MultiModalProcessor(
|
||||
):
|
||||
"""Multimodal processor for HyperCLOVAX V2 (32B Think model)."""
|
||||
|
||||
def get_dummy_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
) -> ProcessorInputs:
|
||||
num_images = mm_counts.get("image", 0)
|
||||
num_videos = mm_counts.get("video", 0)
|
||||
prompt_text = V2_IMAGE_TOKEN * num_images + V2_VIDEO_TOKEN * num_videos
|
||||
|
||||
builder = self.dummy_inputs
|
||||
dummy_mm_data = builder.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data, validate=False)
|
||||
|
||||
tokenizer = self.info.get_tokenizer()
|
||||
prompt = tokenizer.encode(
|
||||
prompt_text,
|
||||
**self.info.default_tok_params.get_encode_kwargs(),
|
||||
)
|
||||
|
||||
return ProcessorInputs(prompt=prompt, mm_data_items=dummy_mm_items)
|
||||
|
||||
def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str:
|
||||
return self.dummy_inputs.get_dummy_text(mm_counts)
|
||||
|
||||
@@ -267,7 +257,7 @@ class HCXVisionV2MultiModalProcessor(
|
||||
hf_config = self.info.get_hf_config()
|
||||
|
||||
# Use token IDs directly from config.
|
||||
# This matches what get_dummy_processor_inputs uses, ensuring consistency.
|
||||
# This matches what get_dummy_inputs uses, ensuring consistency.
|
||||
placeholder: dict[str, int] = {
|
||||
"image": hf_config.image_token_id, # 128060 for <|IMAGE_PAD|>
|
||||
"video": hf_config.video_token_id, # 128061 for <|VIDEO_PAD|>
|
||||
|
||||
@@ -176,29 +176,6 @@ class KimiAudioDummyInputsBuilder(BaseDummyInputsBuilder[KimiAudioProcessingInfo
|
||||
),
|
||||
}
|
||||
|
||||
def get_dummy_processor_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
) -> ProcessorInputs:
|
||||
dummy_mm_data = self.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data)
|
||||
|
||||
num_audios = mm_counts.get("audio", 0)
|
||||
dummy_tokens = (
|
||||
[198]
|
||||
if num_audios == 0
|
||||
else [
|
||||
KimiAudioProcessor.KIMIA_MEDIA_BEGIN,
|
||||
KimiAudioProcessor.KIMIA_TEXT_BLANK,
|
||||
KimiAudioProcessor.KIMIA_MEDIA_END,
|
||||
]
|
||||
* num_audios
|
||||
)
|
||||
|
||||
return ProcessorInputs(prompt=dummy_tokens, mm_data_items=dummy_mm_items)
|
||||
|
||||
|
||||
# Field config for Kimi-Audio multimodal data
|
||||
_KIMIAUDIO_FIELD_CONFIG = {
|
||||
@@ -228,6 +205,30 @@ class KimiAudioMultiModalDataParser(MultiModalDataParser):
|
||||
class KimiAudioMultiModalProcessor(BaseMultiModalProcessor[KimiAudioProcessingInfo]):
|
||||
"""vLLM multi-modal processor wrapper for Kimi-Audio."""
|
||||
|
||||
def get_dummy_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
) -> ProcessorInputs:
|
||||
builder = self.dummy_inputs
|
||||
dummy_mm_data = builder.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data)
|
||||
|
||||
num_audios = mm_counts.get("audio", 0)
|
||||
dummy_tokens = (
|
||||
[198]
|
||||
if num_audios == 0
|
||||
else [
|
||||
KimiAudioProcessor.KIMIA_MEDIA_BEGIN,
|
||||
KimiAudioProcessor.KIMIA_TEXT_BLANK,
|
||||
KimiAudioProcessor.KIMIA_MEDIA_END,
|
||||
]
|
||||
* num_audios
|
||||
)
|
||||
|
||||
return ProcessorInputs(prompt=dummy_tokens, mm_data_items=dummy_mm_items)
|
||||
|
||||
def _get_hf_mm_inputs(
|
||||
self,
|
||||
mm_items: MultiModalDataItems,
|
||||
|
||||
@@ -438,31 +438,38 @@ class MossTranscribeDiarizeDummyInputsBuilder(
|
||||
)
|
||||
}
|
||||
|
||||
def get_dummy_processor_inputs(
|
||||
|
||||
class MossTranscribeDiarizeMultiModalProcessor(
|
||||
BaseMultiModalProcessor[MossTranscribeDiarizeProcessingInfo]
|
||||
):
|
||||
def get_dummy_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
) -> ProcessorInputs:
|
||||
dummy_mm_data = self.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data)
|
||||
num_audios = mm_counts.get("audio", 0)
|
||||
builder = self.dummy_inputs
|
||||
tokenizer = self.info.get_tokenizer()
|
||||
prompt = cached_encode(
|
||||
tokenizer,
|
||||
AUDIO_PLACEHOLDER * num_audios,
|
||||
add_special_tokens=False,
|
||||
) or cached_encode(
|
||||
tokenizer,
|
||||
"\n",
|
||||
add_special_tokens=False,
|
||||
)
|
||||
|
||||
dummy_mm_data = builder.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data)
|
||||
|
||||
num_audios = mm_counts.get("audio", 0)
|
||||
if num_audios > 0:
|
||||
prompt = cached_encode(
|
||||
tokenizer,
|
||||
AUDIO_PLACEHOLDER * num_audios,
|
||||
add_special_tokens=False,
|
||||
)
|
||||
else:
|
||||
prompt = cached_encode(
|
||||
tokenizer,
|
||||
"\n",
|
||||
add_special_tokens=False,
|
||||
)
|
||||
|
||||
return ProcessorInputs(prompt=prompt, mm_data_items=dummy_mm_items)
|
||||
|
||||
|
||||
class MossTranscribeDiarizeMultiModalProcessor(
|
||||
BaseMultiModalProcessor[MossTranscribeDiarizeProcessingInfo]
|
||||
):
|
||||
def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str:
|
||||
return self.dummy_inputs.get_dummy_text(mm_counts)
|
||||
|
||||
|
||||
@@ -221,18 +221,22 @@ class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):
|
||||
)
|
||||
}
|
||||
|
||||
def get_dummy_processor_inputs(
|
||||
|
||||
class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo]):
|
||||
def get_dummy_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
# For test_common.py only
|
||||
mm_data: MultiModalDataDict | None = None,
|
||||
) -> ProcessorInputs:
|
||||
builder = self.dummy_inputs
|
||||
tokenizer = self.info.get_tokenizer()
|
||||
|
||||
dummy_text = self.get_dummy_text(mm_counts)
|
||||
dummy_text = builder.get_dummy_text(mm_counts)
|
||||
dummy_mm_data = (
|
||||
self.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
builder.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
if mm_data is None
|
||||
else mm_data
|
||||
)
|
||||
@@ -256,8 +260,6 @@ class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):
|
||||
|
||||
return ProcessorInputs(prompt=dummy_tokens, mm_data_items=dummy_mm_items)
|
||||
|
||||
|
||||
class PixtralMultiModalProcessor(BaseMultiModalProcessor[PixtralProcessingInfo]):
|
||||
# The tokens are already inserted by the chat template,
|
||||
# so we just double check that they exist
|
||||
def _maybe_apply_prompt_updates(
|
||||
|
||||
@@ -149,19 +149,23 @@ class VoxtralDummyInputsBuilder(BaseDummyInputsBuilder[VoxtralProcessingInfo]):
|
||||
)
|
||||
}
|
||||
|
||||
def get_dummy_processor_inputs(
|
||||
|
||||
class VoxtralMultiModalProcessor(BaseMultiModalProcessor[VoxtralProcessingInfo]):
|
||||
def get_dummy_inputs(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
mm_options: MultiModalDummyOptions,
|
||||
# For test_common.py only
|
||||
mm_data: MultiModalDataDict | None = None,
|
||||
) -> ProcessorInputs:
|
||||
builder = self.dummy_inputs
|
||||
tokenizer = self.info.get_tokenizer()
|
||||
feature_extractor = self.info.get_feature_extractor()
|
||||
|
||||
dummy_text = self.get_dummy_text(mm_counts)
|
||||
dummy_text = builder.get_dummy_text(mm_counts)
|
||||
dummy_mm_data = (
|
||||
self.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
builder.get_dummy_mm_data(seq_len, mm_counts, mm_options)
|
||||
if mm_data is None
|
||||
else mm_data
|
||||
)
|
||||
@@ -197,8 +201,6 @@ class VoxtralDummyInputsBuilder(BaseDummyInputsBuilder[VoxtralProcessingInfo]):
|
||||
|
||||
return ProcessorInputs(prompt=dummy_tokens, mm_data_items=dummy_mm_items)
|
||||
|
||||
|
||||
class VoxtralMultiModalProcessor(BaseMultiModalProcessor[VoxtralProcessingInfo]):
|
||||
# The tokens are already inserted by the chat template,
|
||||
# so we just double check that they exist
|
||||
def _maybe_apply_prompt_updates(
|
||||
|
||||
Reference in New Issue
Block a user