@tduncan / Nvidia-Optimization-Stack-VLLM / commits / 4a6440acef

Bump Transformers version to 5.13.0 (#47867)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

Harry Mellor committed Jul 11, 2026 at 08:56 UTC 4a6440acefbd4d977620bdb6dfb7fb325cd9bda7
12 files changed +30 -24
requirements/test/cpu.txt
+2 -2
@@ -962,7 +962,7 @@ s3transfer==0.10.3
962 # via boto3
963 sacrebleu==2.4.3
964 # via lm-eval
965 -safetensors==0.7.0
965 +safetensors==0.8.0
966 # via
967 # -r requirements/test/../common.txt
968 # accelerate
@@ -1159,7 +1159,7 @@ tqdm==4.67.3
1159 # segmentation-models-pytorch
1160 # sentence-transformers
1161 # transformers
1162 -transformers==5.10.4
1162 +transformers @ git+https://github.com/huggingface/transformers.git@bef013c4a2180aeef83355bd200a7383ce1ac7e3
1163 # via
1164 # -r requirements/test/../common.txt
1165 # -r requirements/test/cuda.in
requirements/test/cuda.in
+1 -1
@@ -39,7 +39,7 @@ open_clip_torch==2.32.0 # Required for nemotron_vl test, Nemotron Parse in test_
39 datamodel_code_generator # required for minicpm3 test
40 lm-eval[api]>=0.4.12 # required for model evaluation test
41 mteb[bm25s]>=2, <3 # required for mteb test
42 -transformers==5.10.4
42 +transformers @ git+https://github.com/huggingface/transformers.git@v5.13-release
43 tokenizers==0.22.2
44 schemathesis>=4.0.0 # Required for openai schema test.
45 # quantization
requirements/test/cuda.txt
+2 -2
@@ -1053,7 +1053,7 @@ s3transfer==0.10.3
1053 # via boto3
1054 sacrebleu==2.4.3
1055 # via lm-eval
1056 -safetensors==0.7.0
1056 +safetensors==0.8.0
1057 # via
1058 # -c requirements/common.txt
1059 # -r requirements/test/../common.txt
@@ -1261,7 +1261,7 @@ tqdm==4.67.3
1261 # segmentation-models-pytorch
1262 # sentence-transformers
1263 # transformers
1264 -transformers==5.10.4
1264 +transformers @ git+https://github.com/huggingface/transformers.git@bef013c4a2180aeef83355bd200a7383ce1ac7e3
1265 # via
1266 # -c requirements/common.txt
1267 # -r requirements/test/../common.txt
requirements/test/nightly-torch.txt
+1 -1
@@ -29,7 +29,7 @@ opencv-python-headless >= 4.13.0 # required for video test
29 datamodel_code_generator # required for minicpm3 test
30 lm-eval[api]>=0.4.12 # required for model evaluation test
31 mteb[bm25s]>=2, <3 # required for mteb test
32 -transformers==5.10.4
32 +transformers @ git+https://github.com/huggingface/transformers.git@v5.13-release
33 tokenizers==0.22.2
34 schemathesis>=4.0.0 # Required for openai schema test.
35 # quantization
requirements/test/rocm.in
+1 -1
@@ -35,7 +35,7 @@ open_clip_torch==2.32.0 # Required for nemotron_vl test, Nemotron Parse in test_
35 datamodel_code_generator # required for minicpm3 test
36 lm-eval[api]>=0.4.12 # required for model evaluation test
37 mteb[bm25s]>=2, <3 # required for mteb test
38 -transformers==5.10.4
38 +transformers @ git+https://github.com/huggingface/transformers.git@v5.13-release
39 tokenizers==0.22.2
40 schemathesis>=4.0.0 # Required for openai schema test
41 # quantization
requirements/test/rocm.txt
+2 -2
@@ -1035,7 +1035,7 @@ s3transfer==0.16.0
1035 # via boto3
1036 sacrebleu==2.6.0
1037 # via lm-eval
1038 -safetensors==0.7.0
1038 +safetensors==0.8.0
1039 # via
1040 # -c requirements/common.txt
1041 # -r requirements/test/../common.txt
@@ -1218,7 +1218,7 @@ tqdm==4.67.3
1218 # sentence-transformers
1219 # tilelang
1220 # transformers
1221 -transformers==5.10.4
1221 +transformers @ git+https://github.com/huggingface/transformers.git@bef013c4a2180aeef83355bd200a7383ce1ac7e3
1222 # via
1223 # -c requirements/common.txt
1224 # -r requirements/test/../common.txt
requirements/test/xpu.in
+1 -1
@@ -17,7 +17,7 @@ accelerate
17 arctic-inference
18 lm_eval[api]>=0.4.12
19 modelscope<1.38
20 -transformers==5.10.4
20 +transformers @ git+https://github.com/huggingface/transformers.git@v5.13-release
21
22 # --- Audio Processing ---
23 librosa
requirements/test/xpu.txt
+2 -2
@@ -779,7 +779,7 @@ rpds-py==0.30.0
779 # referencing
780 sacrebleu==2.6.0
781 # via lm-eval
782 -safetensors==0.7.0
782 +safetensors==0.8.0
783 # via
784 # -c requirements/common.txt
785 # -r requirements/test/../common.txt
@@ -940,7 +940,7 @@ tqdm==4.67.3
940 # pqdm
941 # sentence-transformers
942 # transformers
943 -transformers==5.10.4
943 +transformers @ git+https://github.com/huggingface/transformers.git@bef013c4a2180aeef83355bd200a7383ce1ac7e3
944 # via
945 # -c requirements/common.txt
946 # -r requirements/test/../common.txt
vllm/model_executor/models/granitemoehybrid.py
+4
@@ -316,8 +316,12 @@ class GraniteMoeHybridAttention(nn.Module):
316
317
318 ALL_DECODER_LAYER_TYPES = {
319 + # Transformers < 5.13.0
320 "attention": GraniteMoeHybridAttentionDecoderLayer,
321 "mamba": GraniteMoeHybridMambaDecoderLayer,
322 + # Transformers >= 5.13.0
323 + "full_attention": GraniteMoeHybridAttentionDecoderLayer,
324 + "linear_attention": GraniteMoeHybridMambaDecoderLayer,
325 }
326
327
vllm/model_executor/models/hunyuan_vision.py
+11 -3
@@ -73,6 +73,7 @@ from vllm.multimodal.processing import (
73 BaseProcessingInfo,
74 PromptReplacement,
75 PromptUpdate,
76 + PromptUpdateDetails,
77 )
78 from vllm.sequence import IntermediateTensors
79 from vllm.transformers_utils.configs.hunyuan_vl import (
@@ -751,8 +752,10 @@ class HunYuanVLMultiModalProcessor(BaseMultiModalProcessor[HunYuanVLProcessingIn
752 hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
753 image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs)
754
754 - placeholder = {
755 + token_ids = {
756 "image": hf_processor.image_token_id,
757 + "image_start": hf_processor.image_start_token_id,
758 + "image_end": hf_processor.image_end_token_id,
759 }
760
761 merge_size = image_processor.merge_size
@@ -766,12 +769,17 @@ class HunYuanVLMultiModalProcessor(BaseMultiModalProcessor[HunYuanVLProcessingIn
769 num_tokens = (int(grid_h) // merge_size) * (
770 int(grid_w) // merge_size + 1
771 ) + 2
769 - return [placeholder[modality]] * num_tokens
772 + tokens = (
773 + [token_ids[f"{modality}_start"]]
774 + + [token_ids[modality]] * num_tokens
775 + + [token_ids[f"{modality}_end"]]
776 + )
777 + return PromptUpdateDetails.select_token_id(tokens, token_ids[modality])
778
779 return [
780 PromptReplacement(
781 modality=modality,
774 - target=[placeholder[modality]],
782 + target=[token_ids[modality]],
783 replacement=partial(get_replacement_hunyuan_vl, modality=modality),
784 )
785 for modality in ("image",)
vllm/model_executor/models/olmo3.py
+3 -5
@@ -137,11 +137,9 @@ class Olmo3Attention(nn.Module):
137 )
138
139 # Rotary embeddings. Rope scaling is only applied on full attention layers.
140 - if sliding_window is None:
141 - rope_parameters = self.config.rope_parameters
142 - else:
143 - rope_theta = self.config.rope_parameters["rope_theta"]
144 - rope_parameters = {"rope_type": "default", "rope_theta": rope_theta}
140 + rope_parameters = self.config.rope_parameters
141 + attn_type = "full_attention" if sliding_window is None else "sliding_attention"
142 + rope_parameters = rope_parameters.get(attn_type, rope_parameters)
143 self.rotary_emb = get_rope(
144 self.head_dim,
145 max_position=self.max_position_embeddings,
vllm/transformers_utils/processors/__init__.py
-4
@@ -21,8 +21,6 @@ __all__ = [
21 "GLM4VProcessor",
22 "Granite4VisionProcessor",
23 "H2OVLProcessor",
24 - "HunYuanVLProcessor",
25 - "HunYuanVLImageProcessor",
24 "Moondream3Processor",
25 "InternVLProcessor",
26 "IsaacProcessor",
@@ -58,8 +56,6 @@ _CLASS_TO_MODULE: dict[str, str] = {
56 "GLM4VProcessor": "vllm.transformers_utils.processors.glm4v",
57 "Granite4VisionProcessor": "vllm.transformers_utils.processors.granite4_vision",
58 "H2OVLProcessor": "vllm.transformers_utils.processors.h2ovl",
61 - "HunYuanVLProcessor": "vllm.transformers_utils.processors.hunyuan_vl",
62 - "HunYuanVLImageProcessor": "vllm.transformers_utils.processors.hunyuan_vl_image",
59 "InternVLProcessor": "vllm.transformers_utils.processors.internvl",
60 "IsaacProcessor": "vllm.transformers_utils.processors.isaac",
61 "KimiAudioProcessor": "vllm.transformers_utils.processors.kimi_audio",