Bump Transformers version to 5.13.0 (#47867)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-07-11 00:56:14 -07:00
committed by GitHub
parent bec0a4ede6
commit 4a6440acef
12 changed files with 30 additions and 24 deletions
@@ -316,8 +316,12 @@ class GraniteMoeHybridAttention(nn.Module):
ALL_DECODER_LAYER_TYPES = {
# Transformers < 5.13.0
"attention": GraniteMoeHybridAttentionDecoderLayer,
"mamba": GraniteMoeHybridMambaDecoderLayer,
# Transformers >= 5.13.0
"full_attention": GraniteMoeHybridAttentionDecoderLayer,
"linear_attention": GraniteMoeHybridMambaDecoderLayer,
}
+11 -3
View File
@@ -73,6 +73,7 @@ from vllm.multimodal.processing import (
BaseProcessingInfo,
PromptReplacement,
PromptUpdate,
PromptUpdateDetails,
)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.hunyuan_vl import (
@@ -751,8 +752,10 @@ class HunYuanVLMultiModalProcessor(BaseMultiModalProcessor[HunYuanVLProcessingIn
hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs)
placeholder = {
token_ids = {
"image": hf_processor.image_token_id,
"image_start": hf_processor.image_start_token_id,
"image_end": hf_processor.image_end_token_id,
}
merge_size = image_processor.merge_size
@@ -766,12 +769,17 @@ class HunYuanVLMultiModalProcessor(BaseMultiModalProcessor[HunYuanVLProcessingIn
num_tokens = (int(grid_h) // merge_size) * (
int(grid_w) // merge_size + 1
) + 2
return [placeholder[modality]] * num_tokens
tokens = (
[token_ids[f"{modality}_start"]]
+ [token_ids[modality]] * num_tokens
+ [token_ids[f"{modality}_end"]]
)
return PromptUpdateDetails.select_token_id(tokens, token_ids[modality])
return [
PromptReplacement(
modality=modality,
target=[placeholder[modality]],
target=[token_ids[modality]],
replacement=partial(get_replacement_hunyuan_vl, modality=modality),
)
for modality in ("image",)
+3 -5
View File
@@ -137,11 +137,9 @@ class Olmo3Attention(nn.Module):
)
# Rotary embeddings. Rope scaling is only applied on full attention layers.
if sliding_window is None:
rope_parameters = self.config.rope_parameters
else:
rope_theta = self.config.rope_parameters["rope_theta"]
rope_parameters = {"rope_type": "default", "rope_theta": rope_theta}
rope_parameters = self.config.rope_parameters
attn_type = "full_attention" if sliding_window is None else "sliding_attention"
rope_parameters = rope_parameters.get(attn_type, rope_parameters)
self.rotary_emb = get_rope(
self.head_dim,
max_position=self.max_position_embeddings,