[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
+16 -11
View File
@@ -54,8 +54,19 @@ from vllm.sequence import IntermediateTensors
logger = init_logger(__name__)
def get_mlp_layer_types(config) -> list[str]:
"""Return which layers are MoE, rejecting configs predating Transformers support."""
mlp_layer_types = getattr(config, "mlp_layer_types", None)
if mlp_layer_types is None:
raise ValueError(
"Laguna requires the `LagunaConfig` from Transformers, which is "
"available from v5.17.0. Please update Transformers to run this model."
)
return mlp_layer_types
class LagunaMLP(nn.Module):
"""Dense MLP for Laguna (used in mlp_only_layers)."""
"""Dense MLP for Laguna."""
def __init__(
self,
@@ -206,7 +217,7 @@ class LagunaMoE(nn.Module):
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
renormalize=True,
quant_config=quant_config,
prefix=f"{prefix}.experts",
scoring_func="sigmoid",
@@ -293,13 +304,12 @@ class LagunaAttention(nn.Module):
else:
self.sliding_window = None
# QKV projection (no bias for Laguna)
self.qkv_proj = QKVParallelLinear(
self.hidden_size,
self.head_dim,
self.total_num_heads,
self.total_num_kv_heads,
bias=config.qkv_bias,
bias=config.attention_bias,
quant_config=quant_config,
prefix=f"{prefix}.qkv_proj",
)
@@ -506,14 +516,9 @@ class LagunaDecoderLayer(nn.Module):
),
)
# Check if this layer uses MoE or dense MLP (matches Qwen2/Qwen3 convention)
mlp_only_layers = (
[] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers
)
self.is_moe_layer = (
(layer_idx not in mlp_only_layers)
and (config.num_experts > 0)
and ((layer_idx + 1) % config.decoder_sparse_step == 0)
get_mlp_layer_types(config)[layer_idx] == "sparse"
and config.num_experts > 0
)
if self.is_moe_layer:
+39 -6
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import importlib.metadata
import math
from collections.abc import Iterable, Mapping, Sequence
from dataclasses import dataclass, fields
@@ -12,15 +13,12 @@ import torch.nn as nn
from mistral_common.protocol.instruct.chunk import ImageChunk, TextChunk
from mistral_common.protocol.instruct.messages import UserMessage
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from packaging.version import Version
from transformers import BatchFeature, PixtralVisionConfig
from transformers.models.pixtral.image_processing_pixtral import (
_num_image_tokens as _get_pixtral_hf_num_image_tokens,
)
from transformers.models.pixtral.modeling_pixtral import (
PixtralRotaryEmbedding,
apply_rotary_pos_emb,
position_ids_in_meshgrid,
)
from transformers.models.pixtral.modeling_pixtral import apply_rotary_pos_emb
from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
@@ -92,6 +90,39 @@ from .vision import (
PATCH_MERGE = "patch_merge"
# Transformers 5.17 renamed Pixtral's rotary embedding and switched it to axial
# RoPE, which takes 2D (height, width) positions instead of flattened grid ids.
TRANSFORMERS_VERSION = importlib.metadata.version("transformers")
TRANSFORMERS_WITH_AXIAL_ROPE = Version(TRANSFORMERS_VERSION) >= Version("5.17.0.dev0")
if TRANSFORMERS_WITH_AXIAL_ROPE:
from transformers.models.pixtral.modeling_pixtral import (
PixtralVisionRotaryEmbedding,
)
else:
from transformers.models.pixtral.modeling_pixtral import (
PixtralRotaryEmbedding as PixtralVisionRotaryEmbedding,
)
from transformers.models.pixtral.modeling_pixtral import (
position_ids_in_meshgrid as flat_position_ids_in_meshgrid,
)
def position_ids_in_meshgrid(
patch_embeds_list: list[torch.Tensor],
max_width: int,
) -> torch.Tensor:
if not TRANSFORMERS_WITH_AXIAL_ROPE:
return flat_position_ids_in_meshgrid(patch_embeds_list, max_width)
positions = []
for patch in patch_embeds_list:
height, width = patch.shape[-2:]
h_ids, w_ids = torch.meshgrid(
torch.arange(height), torch.arange(width), indexing="ij"
)
positions.append(torch.stack([h_ids.flatten(), w_ids.flatten()], dim=-1))
return torch.cat(positions, dim=0)
def _make_packed_sequence_metadata(
sequence_lengths: list[int],
@@ -1467,7 +1498,9 @@ class PixtralHFVisionModel(nn.Module):
self.dtype = next(self.parameters()).dtype
self.device = next(self.parameters()).device
self.patch_positional_embedding = PixtralRotaryEmbedding(config, self.device)
self.patch_positional_embedding = PixtralVisionRotaryEmbedding(config).to(
self.device
)
def forward(
self,