[Refactor] Remove dead code in multiple files (#49745)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -194,7 +194,6 @@ class HunYuanVLTextConfig(PretrainedConfig):
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.rope_scaling = rope_scaling
|
||||
# self._rope_scaling_validation() # TODO: Need validation?
|
||||
self.attention_bias = attention_bias
|
||||
self.attention_dropout = attention_dropout
|
||||
|
||||
@@ -206,46 +205,6 @@ class HunYuanVLTextConfig(PretrainedConfig):
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
def _rope_scaling_validation(self):
|
||||
"""
|
||||
Validate the `rope_scaling` configuration.
|
||||
"""
|
||||
if self.rope_scaling is None:
|
||||
return
|
||||
|
||||
if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
|
||||
raise ValueError(
|
||||
"`rope_scaling` must be a dictionary with with two fields, `type` and "
|
||||
f"`factor` or `type` and `alpha`, got {self.rope_scaling}"
|
||||
)
|
||||
rope_scaling_type = self.rope_scaling.get("type", None)
|
||||
rope_scaling_factor = self.rope_scaling.get("factor", None)
|
||||
rope_scaling_alpha = self.rope_scaling.get("alpha", None)
|
||||
if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:
|
||||
raise ValueError(
|
||||
"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], "
|
||||
f"got {rope_scaling_type}"
|
||||
)
|
||||
if rope_scaling_factor is None and rope_scaling_alpha is None:
|
||||
raise ValueError(
|
||||
"`rope_scaling`'s factor or alpha field must be have one, "
|
||||
"got both of none"
|
||||
)
|
||||
if rope_scaling_factor is not None and (
|
||||
not isinstance(rope_scaling_factor, float) or rope_scaling_factor <= 1.0
|
||||
):
|
||||
raise ValueError(
|
||||
"`rope_scaling`'s factor field must be a float > 1.0, "
|
||||
f"got {rope_scaling_factor}"
|
||||
)
|
||||
if rope_scaling_alpha is not None and (
|
||||
not isinstance(rope_scaling_alpha, float) or rope_scaling_alpha <= 1.0
|
||||
):
|
||||
raise ValueError(
|
||||
"`rope_scaling`'s alpha field must be a float > 1.0, "
|
||||
f"got {rope_scaling_alpha}"
|
||||
)
|
||||
|
||||
|
||||
class HunYuanVLConfig(PretrainedConfig):
|
||||
model_type = "hunyuan_vl"
|
||||
|
||||
@@ -72,19 +72,6 @@ class MiniCPMOProcessor(ProcessorMixin):
|
||||
self.version = getattr(image_processor, "version", None)
|
||||
self.pool_step = pool_step
|
||||
|
||||
def _safe_get_token_id(self, attr_name, default_token_str):
|
||||
"""Get token ID safely, with fallback to default."""
|
||||
val = getattr(self.tokenizer, attr_name, None)
|
||||
if val is None:
|
||||
val = self.tokenizer.convert_tokens_to_ids(default_token_str)
|
||||
if val is None:
|
||||
return -1
|
||||
return val
|
||||
|
||||
def _safe_get_token_str(self, attr_name, default_token_str):
|
||||
"""Get token string safely, with fallback to default."""
|
||||
return getattr(self.tokenizer, attr_name, default_token_str)
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput],
|
||||
|
||||
@@ -15,7 +15,6 @@ feeds decoded frames to the processor.
|
||||
|
||||
import math
|
||||
|
||||
import regex as re
|
||||
import torch
|
||||
from torchvision.transforms import InterpolationMode
|
||||
from transformers import AutoTokenizer, BatchFeature
|
||||
@@ -554,85 +553,6 @@ class MiniMaxVLProcessor(ProcessorMixin):
|
||||
self.VISION_END_TOKEN
|
||||
)
|
||||
|
||||
def _prune_video_tokens(
|
||||
self,
|
||||
input_text: str,
|
||||
video_segments: list[int],
|
||||
video_token: str,
|
||||
) -> str:
|
||||
"""Prune video tokens by temporal_patch_size (e.g., 2:1).
|
||||
|
||||
Expects the prompt to carry exactly sum(video_segments) video tokens
|
||||
— i.e. one token per *sampled* frame — then drops tokens.
|
||||
"""
|
||||
# If no videos or temporal_patch_size <= 1, no pruning needed
|
||||
if not video_segments or self.video_processor.temporal_patch_size <= 1:
|
||||
return input_text
|
||||
|
||||
# Split while keeping delimiters
|
||||
special_tokens = [video_token]
|
||||
pattern = "|".join(map(re.escape, special_tokens))
|
||||
parts = re.split(f"({pattern})", input_text)
|
||||
|
||||
def is_timestamp(text: str) -> bool:
|
||||
"""Check if text ends with timestamp format like ']<]0.0 seconds[>['"""
|
||||
return (
|
||||
text.endswith("seconds[>[")
|
||||
or text.endswith("seconds[>[ ")
|
||||
or text.endswith("seconds [>[")
|
||||
or text.endswith("seconds [>[ ")
|
||||
)
|
||||
|
||||
def extract_timestamp(text: str) -> str:
|
||||
"""Extract timestamp text from the end, starting from ']<]'"""
|
||||
start_index = text.rfind("]<]")
|
||||
if start_index == -1:
|
||||
raise ValueError(f"Failed to extract timestamp: {text}")
|
||||
return text[start_index:]
|
||||
|
||||
# Build new text with pruned video tokens
|
||||
final_parts = []
|
||||
current_seg_idx = 0 # Which video segment we're in
|
||||
frame_in_seg = 0 # Frame index within current segment
|
||||
last_timestamp_len = 0 # Length of timestamp to potentially remove
|
||||
|
||||
for part in parts:
|
||||
if part == video_token:
|
||||
if current_seg_idx < len(video_segments):
|
||||
if frame_in_seg % self.video_processor.temporal_patch_size == 0:
|
||||
# Keep this video token
|
||||
final_parts.append(part)
|
||||
frame_in_seg += 1
|
||||
if frame_in_seg >= video_segments[current_seg_idx]:
|
||||
current_seg_idx += 1
|
||||
frame_in_seg = 0
|
||||
last_timestamp_len = 0
|
||||
else:
|
||||
# Skip this video token
|
||||
frame_in_seg += 1
|
||||
if frame_in_seg >= video_segments[current_seg_idx]:
|
||||
current_seg_idx += 1
|
||||
frame_in_seg = 0
|
||||
# Remove the timestamp that was already appended
|
||||
if last_timestamp_len > 0:
|
||||
assert len(final_parts) > 0
|
||||
final_parts[-1] = final_parts[-1][:-last_timestamp_len]
|
||||
last_timestamp_len = 0
|
||||
else:
|
||||
# No more video segments, keep as is
|
||||
final_parts.append(part)
|
||||
last_timestamp_len = 0
|
||||
else:
|
||||
# Text part
|
||||
final_parts.append(part)
|
||||
# Check if this text ends with a timestamp
|
||||
if is_timestamp(part):
|
||||
last_timestamp_len = len(extract_timestamp(part))
|
||||
else:
|
||||
last_timestamp_len = 0
|
||||
|
||||
return "".join(final_parts)
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
images=None,
|
||||
|
||||
Reference in New Issue
Block a user