[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>
This commit is contained in:
Harry Mellor
2026-09-16 06:34:16 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Claude Opus 5
parent f8b5c11468
commit ec4a3a5370
17 changed files with 148 additions and 171 deletions
+49 -2
View File
@@ -13,6 +13,7 @@ from typing import Any, Literal, TypeAlias
import huggingface_hub
import torch
import transformers.configuration_utils as hf_configuration_utils
from huggingface_hub import constants
from packaging.version import Version
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
@@ -126,7 +127,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
RefinedWebModel="RWConfig", # For tiiuae/falcon-7b(-instruct)
mlp_speculator="MLPSpeculatorConfig",
medusa="MedusaConfig",
mellum="MellumConfig",
midashenglm="MiDashengLMConfig",
minimax_m3_vl="MiniMaxM3Config",
minimax_m3_mtp="MiniMaxM3MTPConfig",
@@ -151,7 +151,6 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
qwen3_5_text="Qwen3_5TextConfig",
qwen3_5_moe="Qwen3_5MoeConfig",
qwen3_5_moe_text="Qwen3_5MoeTextConfig",
laguna="LagunaConfig",
lfm2_moe="Lfm2MoeConfig",
**{"unlimited-ocr": "UnlimitedOCRConfig"},
**{"deepseek_v41": "DeepseekV41Config"},
@@ -163,6 +162,12 @@ _SPECULATIVE_DECODING_CONFIGS: set[str] = {"eagle", "speculators", "medusa"}
_PATCH_HF_VALIDATE_ROPE: set[str] = {"sarvam_mla"}
# Model types whose checkpoints store shared RoPE parameters alongside the
# per-layer-type dicts (e.g. Laguna's `original_max_position_embeddings`).
# Since transformers 5.17, `validate_rope` treats every top-level value of such
# a dict as a layer's parameters and raises on the shared ones.
_PATCH_HF_NESTED_ROPE_VALIDATION: set[str] = {"laguna"}
# Model types whose checkpoints declare `layer_types` entries that upstream
# transformers has not added to `ALLOWED_LAYER_TYPES` yet, so its strict config
# validation rejects them (e.g. GLM-5.2 `glm_moe_dsa` use
@@ -236,6 +241,11 @@ def _patch_hf_transformers_validate_rope():
def patched_validate_rope(self, *args, **kwargs):
ignore_keys_param = kwargs.pop("ignore_keys", None)
original_ignore_keys = self.ignore_keys_at_rope_validation
if ignore_keys_param is not None:
logger.warning(
"validate_rope() was called with the legacy 'ignore_keys' "
"argument; use 'ignore_keys_at_rope_validation' instead"
)
self.ignore_keys_at_rope_validation = original_ignore_keys or ignore_keys_param
result = _original_validate_rope(self, *args, **kwargs)
return result
@@ -244,6 +254,40 @@ def _patch_hf_transformers_validate_rope():
PretrainedConfig.validate_rope = patched_validate_rope
def _patch_hf_transformers_nested_rope_validation() -> None:
"""Drop shared entries sitting alongside a nested ``rope_parameters`` dict.
Transformers validates a dict with any layer-type key by iterating all of
its values, so a shared entry next to the per-layer dicts raises an
``AttributeError``. The per-layer dicts already carry their own defaults by
the time validation runs, so the shared entries can be dropped.
"""
if hasattr(PretrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):
return
_original_validate_rope = PretrainedConfig.validate_rope
@wraps(_original_validate_rope)
def patched_validate_rope(self, *args, **kwargs):
rope_parameters = getattr(self, "rope_parameters", None)
if isinstance(rope_parameters, dict):
layer_types = set(rope_parameters) & set(
hf_configuration_utils.ALLOWED_LAYER_TYPES
)
if shared_keys := set(rope_parameters) - layer_types:
for key in shared_keys:
del rope_parameters[key]
logger.warning(
"Dropping %s from rope_parameters; entries shared across "
"layer types are not supported",
sorted(shared_keys),
)
return _original_validate_rope(self, *args, **kwargs)
patched_validate_rope.__vllm_nested_rope_patched__ = True # type: ignore[attr-defined]
PretrainedConfig.validate_rope = patched_validate_rope
def _patch_hf_transformers_allowed_layer_types(
extra_layer_types: tuple[str, ...],
) -> None:
@@ -304,6 +348,9 @@ class HFConfigParser(ConfigParserBase):
if model_type in _PATCH_HF_VALIDATE_ROPE:
_patch_hf_transformers_validate_rope()
if model_type in _PATCH_HF_NESTED_ROPE_VALIDATION:
_patch_hf_transformers_nested_rope_validation()
if extra_layer_types := _PATCH_HF_ALLOWED_LAYER_TYPES.get(model_type):
_patch_hf_transformers_allowed_layer_types(extra_layer_types)
@@ -55,10 +55,8 @@ _CLASS_TO_MODULE: dict[str, str] = {
# tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
# `FalconConfig` class from the official HuggingFace transformers library.
"RWConfig": "vllm.transformers_utils.configs.falcon",
"LagunaConfig": "vllm.transformers_utils.configs.laguna",
"Lfm2MoeConfig": "vllm.transformers_utils.configs.lfm2_moe",
"MedusaConfig": "vllm.transformers_utils.configs.medusa",
"MellumConfig": "vllm.transformers_utils.configs.mellum",
"MiDashengLMConfig": "vllm.transformers_utils.configs.midashenglm",
"MiniMaxM3Config": "vllm.transformers_utils.configs.minimax_m3",
"MiniMaxM3MTPConfig": "vllm.transformers_utils.configs.minimax_m3",
@@ -151,10 +149,8 @@ __all__ = [
"HyperCLOVAXConfig",
"IsaacConfig",
"RWConfig",
"LagunaConfig",
"Lfm2MoeConfig",
"MedusaConfig",
"MellumConfig",
"MiDashengLMConfig",
"MiniMaxM3Config",
"MiniMaxM3MTPConfig",
-120
View File
@@ -1,120 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
class LagunaConfig(PretrainedConfig):
model_type = "laguna"
keys_to_ignore_at_inference = ["past_key_values"]
base_model_tp_plan = {
"layers.*.self_attn.q_proj": "colwise",
"layers.*.self_attn.k_proj": "colwise",
"layers.*.self_attn.v_proj": "colwise",
"layers.*.self_attn.g_proj": "colwise",
"layers.*.self_attn.o_proj": "rowwise",
"layers.*.mlp.gate_proj": "colwise",
"layers.*.mlp.up_proj": "colwise",
"layers.*.mlp.down_proj": "rowwise",
}
base_model_pp_plan = {
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
"norm": (["hidden_states"], ["hidden_states"]),
}
def __init__(
self,
vocab_size: int = 100352,
hidden_size: int = 2048,
intermediate_size: int = 8192,
num_hidden_layers: int = 40,
num_attention_heads: int = 48,
num_key_value_heads: int = 8,
head_dim: int = 128,
qkv_bias: bool = False,
attention_bias: bool = False,
gating: bool | str = True,
hidden_act: str = "silu",
max_position_embeddings: int = 131072,
initializer_range: float = 0.02,
rms_norm_eps: float = 1e-6,
use_cache: bool = True,
tie_word_embeddings: bool = False,
rope_theta: float = 500000.0,
rope_scaling: dict | None = None,
rope_parameters: dict | None = None,
partial_rotary_factor: float = 1.0,
attention_dropout: float = 0.0,
sliding_window: int | None = None,
layer_types: list[str] | None = None,
swa_attention_sink_enabled: bool = False,
swa_rope_parameters: dict | None = None,
num_attention_heads_per_layer: list[int] | None = None,
num_experts: int = 256,
num_experts_per_tok: int = 8,
moe_intermediate_size: int = 512,
shared_expert_intermediate_size: int = 512,
norm_topk_prob: bool = True,
decoder_sparse_step: int = 1,
mlp_only_layers: list[int] | None = None,
router_aux_loss_coef: float = 0.001,
output_router_logits: bool = False,
moe_routed_scaling_factor: float = 1.0,
moe_apply_router_weight_on_input: bool = False,
**kwargs,
):
if mlp_only_layers is None:
mlp_only_layers = [0]
# Accept either v4-style (rope_theta + rope_scaling) or v5-style
# (rope_parameters). Translate v5 → v4 so downstream code has one path.
if rope_parameters is not None:
rp = dict(rope_parameters)
rope_theta = float(rp.pop("rope_theta", rope_theta))
rt = rp.pop("rope_type", None)
if rt is not None and rt != "default":
rope_scaling = {"rope_type": rt, **rp}
elif rp and rope_scaling is None:
rope_scaling = {"rope_type": "default", **rp}
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.head_dim = head_dim
self.qkv_bias = qkv_bias
self.attention_bias = attention_bias
self.gating = gating
self.hidden_act = hidden_act
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.rope_scaling = rope_scaling
self.partial_rotary_factor = partial_rotary_factor
self.attention_dropout = attention_dropout
self.sliding_window = sliding_window
self.layer_types = layer_types
self.swa_attention_sink_enabled = swa_attention_sink_enabled
self.swa_rope_parameters = swa_rope_parameters
self.num_attention_heads_per_layer = num_attention_heads_per_layer
self.num_experts = num_experts
self.num_experts_per_tok = num_experts_per_tok
self.moe_intermediate_size = moe_intermediate_size
self.shared_expert_intermediate_size = shared_expert_intermediate_size
self.norm_topk_prob = norm_topk_prob
self.decoder_sparse_step = decoder_sparse_step
self.mlp_only_layers = mlp_only_layers
self.router_aux_loss_coef = router_aux_loss_coef
self.output_router_logits = output_router_logits
self.moe_routed_scaling_factor = moe_routed_scaling_factor
self.moe_apply_router_weight_on_input = moe_apply_router_weight_on_input
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
__all__ = ["LagunaConfig"]
@@ -1,7 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import Qwen3MoeConfig
class MellumConfig(Qwen3MoeConfig):
model_type = "mellum"