@tduncan / Nvidia-Optimization-Stack-VLLM / commits / ec4a3a5370

[CI] Bump Transformers version to 5.17.0 (#56108)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Harry Mellor committed Sep 16, 2026 at 14:34 UTC ec4a3a537068db40afbc9374a67da719c8c9b964

17 files changed +148 -171

requirements/test/cpu.txt

+1 -1

@@ -1174,7 +1174,7 @@ tqdm==4.67.3

1174 # segmentation-models-pytorch

1175 # sentence-transformers

1176 # transformers

1177 -transformers==5.16.1

1177 +transformers==5.17.0

1178 # via

1179 # -r requirements/test/../common.txt

1180 # -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==5.16.1

42 +transformers==5.17.0

43 schemathesis>=4.0.0 # Required for openai schema test.

44 # quantization

45 buildkite-test-collector==0.1.9

requirements/test/cuda.txt

+1 -1

@@ -1281,7 +1281,7 @@ tqdm==4.67.3

1281 # segmentation-models-pytorch

1282 # sentence-transformers

1283 # transformers

1284 -transformers==5.16.1

1284 +transformers==5.17.0

1285 # via

1286 # -c requirements/common.txt

1287 # -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==5.16.1

32 +transformers==5.17.0

33 schemathesis>=4.0.0 # Required for openai schema test.

34 # quantization

35 buildkite-test-collector==0.1.9

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==5.16.1

38 +transformers==5.17.0

39 schemathesis>=4.0.0 # Required for openai schema test

40 # quantization

41 bitsandbytes==0.50.0

requirements/test/rocm.txt

+1 -1

@@ -1233,7 +1233,7 @@ tqdm==4.67.3

1233 # sentence-transformers

1234 # tilelang

1235 # transformers

1236 -transformers==5.16.1

1236 +transformers==5.17.0

1237 # via

1238 # -c requirements/common.txt

1239 # -r requirements/test/../common.txt

requirements/test/xpu.in

+1 -1

@@ -19,7 +19,7 @@ lm_eval[api]>=0.4.12

19 modelscope<1.38

20 runai-model-streamer[s3,gcs,azure]>=0.15.7

21 tensorizer==2.10.1

22 -transformers==5.16.1

22 +transformers==5.17.0

24 # --- Audio Processing ---

25 av

requirements/test/xpu.txt

+1 -1

@@ -1081,7 +1081,7 @@ tqdm==4.70.0

1081 # pqdm

1082 # sentence-transformers

1083 # transformers

1084 -transformers==5.16.1

1084 +transformers==5.17.0

1085 # via

1086 # -c requirements/common.txt

1087 # -r requirements/test/../common.txt

vllm/model_executor/models/laguna.py

+16 -11

@@ -54,8 +54,19 @@ from vllm.sequence import IntermediateTensors

54 logger = init_logger(__name__)

57 +def get_mlp_layer_types(config) -> list[str]:

58 + """Return which layers are MoE, rejecting configs predating Transformers support."""

59 + mlp_layer_types = getattr(config, "mlp_layer_types", None)

60 + if mlp_layer_types is None:

61 + raise ValueError(

62 + "Laguna requires the `LagunaConfig` from Transformers, which is "

63 + "available from v5.17.0. Please update Transformers to run this model."

64 + )

65 + return mlp_layer_types

66 +

67 +

68 class LagunaMLP(nn.Module):

58 - """Dense MLP for Laguna (used in mlp_only_layers)."""

69 + """Dense MLP for Laguna."""

71 def __init__(

72 self,

@@ -206,7 +217,7 @@ class LagunaMoE(nn.Module):

217 top_k=config.num_experts_per_tok,

218 hidden_size=config.hidden_size,

219 intermediate_size=config.moe_intermediate_size,

209 - renormalize=config.norm_topk_prob,

220 + renormalize=True,

221 quant_config=quant_config,

222 prefix=f"{prefix}.experts",

223 scoring_func="sigmoid",

@@ -293,13 +304,12 @@ class LagunaAttention(nn.Module):

304 else:

305 self.sliding_window = None

306

296 - # QKV projection (no bias for Laguna)

307 self.qkv_proj = QKVParallelLinear(

308 self.hidden_size,

309 self.head_dim,

310 self.total_num_heads,

311 self.total_num_kv_heads,

302 - bias=config.qkv_bias,

312 + bias=config.attention_bias,

313 quant_config=quant_config,

314 prefix=f"{prefix}.qkv_proj",

315 )

@@ -506,14 +516,9 @@ class LagunaDecoderLayer(nn.Module):

516 ),

517 )

518

509 - # Check if this layer uses MoE or dense MLP (matches Qwen2/Qwen3 convention)

510 - mlp_only_layers = (

511 - [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers

512 - )

519 self.is_moe_layer = (

514 - (layer_idx not in mlp_only_layers)

515 - and (config.num_experts > 0)

516 - and ((layer_idx + 1) % config.decoder_sparse_step == 0)

520 + get_mlp_layer_types(config)[layer_idx] == "sparse"

521 + and config.num_experts > 0

522 )

523

524 if self.is_moe_layer:

vllm/model_executor/models/pixtral.py

+39 -6

@@ -1,6 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

4 +import importlib.metadata

5 import math

6 from collections.abc import Iterable, Mapping, Sequence

7 from dataclasses import dataclass, fields

@@ -12,15 +13,12 @@ import torch.nn as nn

13 from mistral_common.protocol.instruct.chunk import ImageChunk, TextChunk

14 from mistral_common.protocol.instruct.messages import UserMessage

15 from mistral_common.protocol.instruct.request import ChatCompletionRequest

16 +from packaging.version import Version

17 from transformers import BatchFeature, PixtralVisionConfig

18 from transformers.models.pixtral.image_processing_pixtral import (

19 _num_image_tokens as _get_pixtral_hf_num_image_tokens,

20 )

19 -from transformers.models.pixtral.modeling_pixtral import (

20 - PixtralRotaryEmbedding,

21 - apply_rotary_pos_emb,

22 - position_ids_in_meshgrid,

23 -)

21 +from transformers.models.pixtral.modeling_pixtral import apply_rotary_pos_emb

23 from vllm.config import VllmConfig

24 from vllm.config.multimodal import BaseDummyOptions

@@ -92,6 +90,39 @@ from .vision import (

91 PATCH_MERGE = "patch_merge"

93 +# Transformers 5.17 renamed Pixtral's rotary embedding and switched it to axial

94 +# RoPE, which takes 2D (height, width) positions instead of flattened grid ids.

95 +TRANSFORMERS_VERSION = importlib.metadata.version("transformers")

96 +TRANSFORMERS_WITH_AXIAL_ROPE = Version(TRANSFORMERS_VERSION) >= Version("5.17.0.dev0")

97 +

98 +if TRANSFORMERS_WITH_AXIAL_ROPE:

99 + from transformers.models.pixtral.modeling_pixtral import (

100 + PixtralVisionRotaryEmbedding,

101 + )

102 +else:

103 + from transformers.models.pixtral.modeling_pixtral import (

104 + PixtralRotaryEmbedding as PixtralVisionRotaryEmbedding,

105 + )

106 + from transformers.models.pixtral.modeling_pixtral import (

107 + position_ids_in_meshgrid as flat_position_ids_in_meshgrid,

108 + )

109 +

110 +

111 +def position_ids_in_meshgrid(

112 + patch_embeds_list: list[torch.Tensor],

113 + max_width: int,

114 +) -> torch.Tensor:

115 + if not TRANSFORMERS_WITH_AXIAL_ROPE:

116 + return flat_position_ids_in_meshgrid(patch_embeds_list, max_width)

117 + positions = []

118 + for patch in patch_embeds_list:

119 + height, width = patch.shape[-2:]

120 + h_ids, w_ids = torch.meshgrid(

121 + torch.arange(height), torch.arange(width), indexing="ij"

122 + )

123 + positions.append(torch.stack([h_ids.flatten(), w_ids.flatten()], dim=-1))

124 + return torch.cat(positions, dim=0)

125 +

126

127 def _make_packed_sequence_metadata(

128 sequence_lengths: list[int],

@@ -1467,7 +1498,9 @@ class PixtralHFVisionModel(nn.Module):

1498

1499 self.dtype = next(self.parameters()).dtype

1500 self.device = next(self.parameters()).device

1470 - self.patch_positional_embedding = PixtralRotaryEmbedding(config, self.device)

1501 + self.patch_positional_embedding = PixtralVisionRotaryEmbedding(config).to(

1502 + self.device

1503 + )

1504

1505 def forward(

1506 self,

vllm/models/qwen4_exp/amd/model.py

+12 -6

@@ -76,7 +76,7 @@ from vllm.transformers_utils.configs.qwen4_exp import (

76 from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum

77 from vllm.v1.kv_cache_interface import MambaSpec

79 -from ..config import Qwen4ExpConfig

79 +from ..config import ATTENTION_LAYER_TYPES, QSA_LAYER_TYPE, Qwen4ExpConfig

80 from .hyperconnection import GatedResidual, HyperConnectionConfig

81 from .low_latency_gemm import enable_qwen4_exp_low_latency_gemm

82 from .ple_layer import Qwen4ExpPLELayer

@@ -214,8 +214,11 @@ class Qwen4ExpDecoderLayer(nn.Module):

214 prefix=f"{prefix}.linear_attn",

215 gqa_interleaved_layout=False,

216 )

217 - elif layer_type == "full_attention":

218 - use_qsa = getattr(config, "indexer_n_heads", None) is not None

217 + elif layer_type in ATTENTION_LAYER_TYPES:

218 + use_qsa = (

219 + layer_type == QSA_LAYER_TYPE

220 + or getattr(config, "indexer_n_heads", None) is not None

221 + )

222 if not use_qsa:

223 self.self_attn = Qwen3NextAttention(

224 config,

@@ -311,7 +314,7 @@ class Qwen4ExpDecoderLayer(nn.Module):

314

315 if self.layer_type == "linear_attention":

316 attn_out = self.linear_attn(hidden_states=block_input)

314 - elif self.layer_type == "full_attention":

317 + elif self.layer_type in ATTENTION_LAYER_TYPES:

318 attn_out = self.self_attn(

319 hidden_states=block_input,

320 positions=positions,

@@ -402,8 +405,11 @@ class Qwen4ExpModel(nn.Module):

405 self._qsa_layer_ids = frozenset(

406 layer_idx

407 for layer_idx, layer_type in enumerate(config.layer_types)

405 - if layer_type == "full_attention"

406 - and getattr(config, "indexer_n_heads", None) is not None

408 + if layer_type == QSA_LAYER_TYPE

409 + or (

410 + layer_type == "full_attention"

411 + and getattr(config, "indexer_n_heads", None) is not None

412 + )

413 )

414 self.embed_tokens = VocabParallelEmbedding(self.vocab_size, config.hidden_size)

415

vllm/models/qwen4_exp/config.py

+12 -1

@@ -19,6 +19,12 @@ _QSA_CONFIG_FIELDS = (

19 "indexer_compress_ratio",

20 )

22 +# Transformers labels Qwen4Exp's sparse-attention layers `qwen_sparse_attention`.

23 +# Older checkpoints label them `full_attention` and mark QSA by setting the

24 +# indexer fields for the whole model.

25 +QSA_LAYER_TYPE = "qwen_sparse_attention"

26 +ATTENTION_LAYER_TYPES = ("full_attention", QSA_LAYER_TYPE)

27 +

29 class Qwen4ExpVisionConfig(Qwen3VLVisionConfig):

30 model_type = "qwen4_exp"

@@ -134,6 +140,11 @@ class Qwen4ExpTextConfig(Qwen3NextConfig):

140 def _validate_qsa_config(self) -> None:

141 configured = {name: getattr(self, name, None) for name in _QSA_CONFIG_FIELDS}

142 if all(value is None for value in configured.values()):

143 + if QSA_LAYER_TYPE in (self.layer_types or ()):

144 + raise ValueError(

145 + f"{QSA_LAYER_TYPE} layers require the QSA config fields: "

146 + f"{list(_QSA_CONFIG_FIELDS)}"

147 + )

148 return

149

150 missing = [name for name, value in configured.items() if value is None]

@@ -165,7 +176,7 @@ class Qwen4ExpTextConfig(Qwen3NextConfig):

176 @property

177 def layers_block_type(self) -> list[str]:

178 return [

168 - "attention" if layer_type == "full_attention" else layer_type

179 + "attention" if layer_type in ATTENTION_LAYER_TYPES else layer_type

180 for layer_type in self.layer_types

181 ]

182

vllm/models/qwen4_exp/nvidia/model.py

+12 -6

@@ -75,7 +75,7 @@ from vllm.transformers_utils.configs.qwen4_exp import (

75 from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum

76 from vllm.v1.kv_cache_interface import MambaSpec

78 -from ..config import Qwen4ExpConfig

78 +from ..config import ATTENTION_LAYER_TYPES, QSA_LAYER_TYPE, Qwen4ExpConfig

79 from .hyperconnection import GatedResidual, HyperConnectionConfig

80 from .low_latency_gemm import enable_qwen4_exp_low_latency_gemm

81 from .ple_layer import Qwen4ExpPLELayer

@@ -215,8 +215,11 @@ class Qwen4ExpDecoderLayer(nn.Module):

215 prefix=f"{prefix}.linear_attn",

216 gqa_interleaved_layout=False,

217 )

218 - elif layer_type == "full_attention":

219 - use_qsa = getattr(config, "indexer_n_heads", None) is not None

218 + elif layer_type in ATTENTION_LAYER_TYPES:

219 + use_qsa = (

220 + layer_type == QSA_LAYER_TYPE

221 + or getattr(config, "indexer_n_heads", None) is not None

222 + )

223 if not use_qsa:

224 self.self_attn = Qwen3NextAttention(

225 config,

@@ -314,7 +317,7 @@ class Qwen4ExpDecoderLayer(nn.Module):

317

318 if self.layer_type == "linear_attention":

319 attn_out = self.linear_attn(hidden_states=block_input)

317 - elif self.layer_type == "full_attention":

320 + elif self.layer_type in ATTENTION_LAYER_TYPES:

321 attn_out = self.self_attn(

322 hidden_states=block_input,

323 positions=positions,

@@ -394,8 +397,11 @@ class Qwen4ExpModel(nn.Module):

397 self._qsa_layer_ids = frozenset(

398 layer_idx

399 for layer_idx, layer_type in enumerate(config.layer_types)

397 - if layer_type == "full_attention"

398 - and getattr(config, "indexer_n_heads", None) is not None

400 + if layer_type == QSA_LAYER_TYPE

401 + or (

402 + layer_type == "full_attention"

403 + and getattr(config, "indexer_n_heads", None) is not None

404 + )

405 )

406 self.embed_tokens = VocabParallelEmbedding(self.vocab_size, config.hidden_size)

407

vllm/transformers_utils/config.py

+49 -2

@@ -13,6 +13,7 @@ from typing import Any, Literal, TypeAlias

14 import huggingface_hub

15 import torch

16 +import transformers.configuration_utils as hf_configuration_utils

17 from huggingface_hub import constants

18 from packaging.version import Version

19 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

@@ -126,7 +127,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(

127 RefinedWebModel="RWConfig", # For tiiuae/falcon-7b(-instruct)

128 mlp_speculator="MLPSpeculatorConfig",

129 medusa="MedusaConfig",

129 - mellum="MellumConfig",

130 midashenglm="MiDashengLMConfig",

131 minimax_m3_vl="MiniMaxM3Config",

132 minimax_m3_mtp="MiniMaxM3MTPConfig",

@@ -151,7 +151,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(

151 qwen3_5_text="Qwen3_5TextConfig",

152 qwen3_5_moe="Qwen3_5MoeConfig",

153 qwen3_5_moe_text="Qwen3_5MoeTextConfig",

154 - laguna="LagunaConfig",

154 lfm2_moe="Lfm2MoeConfig",

155 **{"unlimited-ocr": "UnlimitedOCRConfig"},

156 **{"deepseek_v41": "DeepseekV41Config"},

@@ -163,6 +162,12 @@ _SPECULATIVE_DECODING_CONFIGS: set[str] = {"eagle", "speculators", "medusa"}

162

163 _PATCH_HF_VALIDATE_ROPE: set[str] = {"sarvam_mla"}

164

165 +# Model types whose checkpoints store shared RoPE parameters alongside the

166 +# per-layer-type dicts (e.g. Laguna's `original_max_position_embeddings`).

167 +# Since transformers 5.17, `validate_rope` treats every top-level value of such

168 +# a dict as a layer's parameters and raises on the shared ones.

169 +_PATCH_HF_NESTED_ROPE_VALIDATION: set[str] = {"laguna"}

170 +

171 # Model types whose checkpoints declare `layer_types` entries that upstream

172 # transformers has not added to `ALLOWED_LAYER_TYPES` yet, so its strict config

173 # validation rejects them (e.g. GLM-5.2 `glm_moe_dsa` use

@@ -236,6 +241,11 @@ def _patch_hf_transformers_validate_rope():

241 def patched_validate_rope(self, *args, **kwargs):

242 ignore_keys_param = kwargs.pop("ignore_keys", None)

243 original_ignore_keys = self.ignore_keys_at_rope_validation

244 + if ignore_keys_param is not None:

245 + logger.warning(

246 + "validate_rope() was called with the legacy 'ignore_keys' "

247 + "argument; use 'ignore_keys_at_rope_validation' instead"

248 + )

249 self.ignore_keys_at_rope_validation = original_ignore_keys or ignore_keys_param

250 result = _original_validate_rope(self, *args, **kwargs)

251 return result

@@ -244,6 +254,40 @@ def _patch_hf_transformers_validate_rope():

254 PretrainedConfig.validate_rope = patched_validate_rope

255

256

257 +def _patch_hf_transformers_nested_rope_validation() -> None:

258 + """Drop shared entries sitting alongside a nested ``rope_parameters`` dict.

259 +

260 + Transformers validates a dict with any layer-type key by iterating all of

261 + its values, so a shared entry next to the per-layer dicts raises an

262 + ``AttributeError``. The per-layer dicts already carry their own defaults by

263 + the time validation runs, so the shared entries can be dropped.

264 + """

265 + if hasattr(PretrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):

266 + return

267 +

268 + _original_validate_rope = PretrainedConfig.validate_rope

269 +

270 + @wraps(_original_validate_rope)

271 + def patched_validate_rope(self, *args, **kwargs):

272 + rope_parameters = getattr(self, "rope_parameters", None)

273 + if isinstance(rope_parameters, dict):

274 + layer_types = set(rope_parameters) & set(

275 + hf_configuration_utils.ALLOWED_LAYER_TYPES

276 + )

277 + if shared_keys := set(rope_parameters) - layer_types:

278 + for key in shared_keys:

279 + del rope_parameters[key]

280 + logger.warning(

281 + "Dropping %s from rope_parameters; entries shared across "

282 + "layer types are not supported",

283 + sorted(shared_keys),

284 + )

285 + return _original_validate_rope(self, *args, **kwargs)

286 +

287 + patched_validate_rope.__vllm_nested_rope_patched__ = True # type: ignore[attr-defined]

288 + PretrainedConfig.validate_rope = patched_validate_rope

289 +

290 +

291 def _patch_hf_transformers_allowed_layer_types(

292 extra_layer_types: tuple[str, ...],

293 ) -> None:

@@ -304,6 +348,9 @@ class HFConfigParser(ConfigParserBase):

348 if model_type in _PATCH_HF_VALIDATE_ROPE:

349 _patch_hf_transformers_validate_rope()

350

351 + if model_type in _PATCH_HF_NESTED_ROPE_VALIDATION:

352 + _patch_hf_transformers_nested_rope_validation()

353 +

354 if extra_layer_types := _PATCH_HF_ALLOWED_LAYER_TYPES.get(model_type):

355 _patch_hf_transformers_allowed_layer_types(extra_layer_types)

356

vllm/transformers_utils/configs/__init__.py

-4

@@ -55,10 +55,8 @@ _CLASS_TO_MODULE: dict[str, str] = {

55 # tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the

56 # `FalconConfig` class from the official HuggingFace transformers library.

57 "RWConfig": "vllm.transformers_utils.configs.falcon",

58 - "LagunaConfig": "vllm.transformers_utils.configs.laguna",

58 "Lfm2MoeConfig": "vllm.transformers_utils.configs.lfm2_moe",

59 "MedusaConfig": "vllm.transformers_utils.configs.medusa",

61 - "MellumConfig": "vllm.transformers_utils.configs.mellum",

60 "MiDashengLMConfig": "vllm.transformers_utils.configs.midashenglm",

61 "MiniMaxM3Config": "vllm.transformers_utils.configs.minimax_m3",

62 "MiniMaxM3MTPConfig": "vllm.transformers_utils.configs.minimax_m3",

@@ -151,10 +149,8 @@ __all__ = [

149 "HyperCLOVAXConfig",

150 "IsaacConfig",

151 "RWConfig",

154 - "LagunaConfig",

152 "Lfm2MoeConfig",

153 "MedusaConfig",

157 - "MellumConfig",

154 "MiDashengLMConfig",

155 "MiniMaxM3Config",

156 "MiniMaxM3MTPConfig",

vllm/transformers_utils/configs/laguna.py deleted

-120

@@ -1,120 +0,0 @@

1 -# SPDX-License-Identifier: Apache-2.0

2 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 -

4 -from transformers.configuration_utils import PretrainedConfig

5 -

6 -

7 -class LagunaConfig(PretrainedConfig):

8 - model_type = "laguna"

9 - keys_to_ignore_at_inference = ["past_key_values"]

10 - base_model_tp_plan = {

11 - "layers.*.self_attn.q_proj": "colwise",

12 - "layers.*.self_attn.k_proj": "colwise",

13 - "layers.*.self_attn.v_proj": "colwise",

14 - "layers.*.self_attn.g_proj": "colwise",

15 - "layers.*.self_attn.o_proj": "rowwise",

16 - "layers.*.mlp.gate_proj": "colwise",

17 - "layers.*.mlp.up_proj": "colwise",

18 - "layers.*.mlp.down_proj": "rowwise",

19 - }

20 - base_model_pp_plan = {

21 - "embed_tokens": (["input_ids"], ["inputs_embeds"]),

22 - "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),

23 - "norm": (["hidden_states"], ["hidden_states"]),

24 - }

25 -

26 - def __init__(

27 - self,

28 - vocab_size: int = 100352,

29 - hidden_size: int = 2048,

30 - intermediate_size: int = 8192,

31 - num_hidden_layers: int = 40,

32 - num_attention_heads: int = 48,

33 - num_key_value_heads: int = 8,

34 - head_dim: int = 128,

35 - qkv_bias: bool = False,

36 - attention_bias: bool = False,

37 - gating: bool | str = True,

38 - hidden_act: str = "silu",

39 - max_position_embeddings: int = 131072,

40 - initializer_range: float = 0.02,

41 - rms_norm_eps: float = 1e-6,

42 - use_cache: bool = True,

43 - tie_word_embeddings: bool = False,

44 - rope_theta: float = 500000.0,

45 - rope_scaling: dict | None = None,

46 - rope_parameters: dict | None = None,

47 - partial_rotary_factor: float = 1.0,

48 - attention_dropout: float = 0.0,

49 - sliding_window: int | None = None,

50 - layer_types: list[str] | None = None,

51 - swa_attention_sink_enabled: bool = False,

52 - swa_rope_parameters: dict | None = None,

53 - num_attention_heads_per_layer: list[int] | None = None,

54 - num_experts: int = 256,

55 - num_experts_per_tok: int = 8,

56 - moe_intermediate_size: int = 512,

57 - shared_expert_intermediate_size: int = 512,

58 - norm_topk_prob: bool = True,

59 - decoder_sparse_step: int = 1,

60 - mlp_only_layers: list[int] | None = None,

61 - router_aux_loss_coef: float = 0.001,

62 - output_router_logits: bool = False,

63 - moe_routed_scaling_factor: float = 1.0,

64 - moe_apply_router_weight_on_input: bool = False,

65 - **kwargs,

66 - ):

67 - if mlp_only_layers is None:

68 - mlp_only_layers = [0]

69 -

70 - # Accept either v4-style (rope_theta + rope_scaling) or v5-style

71 - # (rope_parameters). Translate v5 → v4 so downstream code has one path.

72 - if rope_parameters is not None:

73 - rp = dict(rope_parameters)

74 - rope_theta = float(rp.pop("rope_theta", rope_theta))

75 - rt = rp.pop("rope_type", None)

76 - if rt is not None and rt != "default":

77 - rope_scaling = {"rope_type": rt, **rp}

78 - elif rp and rope_scaling is None:

79 - rope_scaling = {"rope_type": "default", **rp}

80 -

81 - self.vocab_size = vocab_size

82 - self.hidden_size = hidden_size

83 - self.intermediate_size = intermediate_size

84 - self.num_hidden_layers = num_hidden_layers

85 - self.num_attention_heads = num_attention_heads

86 - self.num_key_value_heads = num_key_value_heads

87 - self.head_dim = head_dim

88 - self.qkv_bias = qkv_bias

89 - self.attention_bias = attention_bias

90 - self.gating = gating

91 - self.hidden_act = hidden_act

92 - self.max_position_embeddings = max_position_embeddings

93 - self.initializer_range = initializer_range

94 - self.rms_norm_eps = rms_norm_eps

95 - self.use_cache = use_cache

96 - self.rope_theta = rope_theta

97 - self.rope_scaling = rope_scaling

98 - self.partial_rotary_factor = partial_rotary_factor

99 - self.attention_dropout = attention_dropout

100 - self.sliding_window = sliding_window

101 - self.layer_types = layer_types

102 - self.swa_attention_sink_enabled = swa_attention_sink_enabled

103 - self.swa_rope_parameters = swa_rope_parameters

104 - self.num_attention_heads_per_layer = num_attention_heads_per_layer

105 - self.num_experts = num_experts

106 - self.num_experts_per_tok = num_experts_per_tok

107 - self.moe_intermediate_size = moe_intermediate_size

108 - self.shared_expert_intermediate_size = shared_expert_intermediate_size

109 - self.norm_topk_prob = norm_topk_prob

110 - self.decoder_sparse_step = decoder_sparse_step

111 - self.mlp_only_layers = mlp_only_layers

112 - self.router_aux_loss_coef = router_aux_loss_coef

113 - self.output_router_logits = output_router_logits

114 - self.moe_routed_scaling_factor = moe_routed_scaling_factor

115 - self.moe_apply_router_weight_on_input = moe_apply_router_weight_on_input

116 -

117 - super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

118 -

119 -

120 -__all__ = ["LagunaConfig"]

vllm/transformers_utils/configs/mellum.py deleted

-7

@@ -1,7 +0,0 @@

1 -# SPDX-License-Identifier: Apache-2.0

2 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 -from transformers import Qwen3MoeConfig

4 -

5 -

6 -class MellumConfig(Qwen3MoeConfig):

7 - model_type = "mellum"