Merge remote-tracking branch 'upstream/main' into patches

This commit is contained in:
2026-09-28 12:06:52 +07:00
170 changed files with 7605 additions and 1142 deletions
+16 -7
View File
@@ -30,6 +30,9 @@ from vllm.model_executor.layers.linear import (
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb
from vllm.model_executor.layers.rotary_embedding.packed_qk_rope import (
packed_qk_rope_,
)
from vllm.model_executor.models.utils import maybe_prefix
from vllm.model_executor.models.vision import (
is_vit_use_data_parallel,
@@ -37,6 +40,7 @@ from vllm.model_executor.models.vision import (
)
from vllm.platforms import current_platform
from vllm.transformers_utils.configs.kimi_k25 import KimiK25VisionConfig
from vllm.triton_utils import HAS_TRITON
from vllm.utils.torch_utils import async_tensor_h2d
logger = init_logger(__name__)
@@ -451,14 +455,19 @@ class MoonViTEncoderLayer(nn.Module):
)
# xqkv: (seqlen, 3, nheads, headdim)
xqkv = xqkv.view(*qkv_shape)
xq, xk, xv = torch.unbind(xqkv, dim=-3)
_apply_rope_input_validation(xq, rope_freqs_cis)
_apply_rope_input_validation(xk, rope_freqs_cis)
rope_cos = rope_freqs_cis.real.contiguous()
rope_sin = rope_freqs_cis.imag.contiguous()
xq = self.apply_rotary_emb(xq, rope_cos, rope_sin)
xk = self.apply_rotary_emb(xk, rope_cos, rope_sin)
# xq/xk alias xqkv, so the in-place rotation below is visible through
# them; only the fallback needs to rebind.
xq, xk, xv = torch.unbind(xqkv, dim=-3)
if HAS_TRITON:
packed_qk_rope_(xqkv, rope_freqs_cis)
else:
_apply_rope_input_validation(xq, rope_freqs_cis)
_apply_rope_input_validation(xk, rope_freqs_cis)
rope_cos = rope_freqs_cis.real.contiguous()
rope_sin = rope_freqs_cis.imag.contiguous()
xq = self.apply_rotary_emb(xq, rope_cos, rope_sin)
xk = self.apply_rotary_emb(xk, rope_cos, rope_sin)
if max_seqlen is None:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+12 -9
View File
@@ -1055,15 +1055,18 @@ class OpenPanguVLForConditionalGeneration(
input_ids: torch.Tensor,
multimodal_embeddings: MultiModalEmbeddings | None = None,
) -> torch.Tensor:
inputs_embeds = self.language_model.embed_input_ids(input_ids)
if multimodal_embeddings is not None:
inputs_embeds = self.embed_input_ids( # type: ignore[call-overload]
input_ids,
inputs_embeds,
multimodal_embeddings,
[self.config.image_token_id, self.config.video_token_id],
)
return inputs_embeds
if multimodal_embeddings is None:
return self.language_model.embed_input_ids(input_ids)
mm_token_ids = input_ids.new_tensor(
[self.config.image_token_id, self.config.video_token_id]
)
is_multimodal = torch.isin(input_ids, mm_token_ids)
return self.embed_input_ids(
input_ids,
multimodal_embeddings,
is_multimodal=is_multimodal,
)
def _process_image_input(
self, image_input: OpenPanguVLImageInputs
+1 -1
View File
@@ -1297,7 +1297,7 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> None:
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
loader = AutoWeightsLoader(self)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+5 -3
View File
@@ -41,7 +41,7 @@ from vllm.model_executor.models.utils import (
from vllm.sequence import IntermediateTensors
from .bert_with_rope import BertWithRope, JinaRobertaModel
from .interfaces import SupportsCrossEncoding
from .interfaces import SupportsCrossEncoding, SupportsLoRA
from .interfaces_base import attn_type, default_pooling_type
@@ -266,7 +266,7 @@ class BgeM3EmbeddingModel(RobertaEmbeddingModel):
@default_pooling_type(seq_pooling_type="CLS")
class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding):
class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding, SupportsLoRA):
"""A model that uses Roberta to provide embedding functionalities.
This class encapsulates the BertModel and provides an interface for
@@ -279,6 +279,7 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding):
"""
is_pooling_model = True
packed_modules_mapping = {"qkv_proj": ["query", "key", "value"]}
jina_to_vllm_mapper = WeightsMapper(
orig_to_new_substr={
"emb_ln": "embeddings.LayerNorm",
@@ -295,7 +296,8 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
self.padding_idx: int = vllm_config.model_config.hf_config.pad_token_id
self.config = config
self.padding_idx: int = config.pad_token_id
self.num_labels = config.num_labels
self.roberta = BertModel(
+13 -8
View File
@@ -282,10 +282,14 @@ class UltravoxMultiModalProcessor(BaseMultiModalProcessor[UltravoxProcessingInfo
def get_replacement_ultravox(item_idx: int):
if "audio_num_tokens" in out_mm_data:
return [replacement_id] * int(out_mm_data["audio_num_tokens"][item_idx])
audio_num_tokens = out_mm_data["audio_num_tokens"]
assert isinstance(audio_num_tokens, torch.Tensor)
return [replacement_id] * int(audio_num_tokens[item_idx])
start = chunks_start_idx[item_idx]
end = chunks_start_idx[item_idx + 1]
audio_token_len = out_mm_data["audio_token_len"][start:end].sum()
audio_token_lens = out_mm_data["audio_token_len"]
assert isinstance(audio_token_lens, torch.Tensor)
audio_token_len = audio_token_lens[start:end].sum()
return [replacement_id] * int(audio_token_len) # type: ignore
return [
@@ -566,7 +570,8 @@ class UltravoxWhisperEncoder(WhisperEncoder):
) -> torch.Tensor:
return (input_lengths - 1) // 2 + 1
def forward(
# Ultravox requires chunk lengths in addition to Whisper's input features.
def forward( # type: ignore[override]
self,
input_features: torch.Tensor,
audio_lens: torch.Tensor,
@@ -960,12 +965,12 @@ def pad_and_concat_to_dim3(
return features
features = [pad_and_concat_to_dim3(f) for f in features]
tensors = [pad_and_concat_to_dim3(f) for f in features]
max_len = max(f.shape[-1] for f in features)
max_len = max(f.shape[-1] for f in tensors)
# Ensure all features have dim=3
features = [f.view(-1, *f.shape[-2:]) for f in features]
tensors = [f.view(-1, *f.shape[-2:]) for f in tensors]
# Pad and concatenate:
# [[B1, 80, M1], [B2, 80, M2]] -> [B1+B2, 80, max(M1, M2)]
features = [F.pad(f, (0, max_len - f.shape[-1])) for f in features]
return torch.cat(features)
tensors = [F.pad(f, (0, max_len - f.shape[-1])) for f in tensors]
return torch.cat(tensors)
@@ -199,6 +199,7 @@ class UnlimitedOCRMultiModalProcessor(DeepseekOCRMultiModalProcessor):
if isinstance(images, ImageEmbeddingItems):
num_image_tokens = images.get_feature_size(item_idx)
else:
assert isinstance(images, ImageProcessorItems)
size = images.get_image_size(item_idx)
# Disable crop mode for multi-image input.
+8 -7
View File
@@ -35,7 +35,6 @@ if TYPE_CHECKING:
from transformers import PreTrainedConfig
from transformers.conversion_mapping import WeightRenaming
from vllm.config.model import ModelConfig
from vllm.model_executor.layers.quantization import QuantizationConfig
logger = init_logger(__name__)
@@ -337,7 +336,7 @@ class AutoWeightsLoader:
"""
# Add persistent registered buffers.
# Non-persistent buffers are excluded, matching PyTorch state_dict().
non_persistent = getattr(module, "_non_persistent_buffers_set", set())
non_persistent: set[str] = getattr(module, "_non_persistent_buffers_set", set())
for buf_name, buf in module.named_buffers(recurse=False):
if buf_name not in child_params and buf_name not in non_persistent:
child_params[buf_name] = buf
@@ -558,7 +557,7 @@ def maybe_fuse_shared_experts(
def get_spec_layer_idx_from_weight_name(
config: "ModelConfig", weight_name: str
config: "PreTrainedConfig", weight_name: str
) -> int | None:
"""Return the MTP layer index a weight belongs to, or None.
@@ -587,7 +586,7 @@ def get_spec_layer_idx_from_weight_name(
def skip_spec_layers(
weights: Iterable[tuple[str, torch.Tensor]], config: "ModelConfig"
weights: Iterable[tuple[str, torch.Tensor]], config: "PreTrainedConfig"
) -> Iterable[tuple[str, torch.Tensor]]:
"""Drop MTP spec-layer weights (loaded by the MTP head, not the base model).
@@ -629,11 +628,11 @@ def init_vllm_registered_model(
@overload
def flatten_bn(x: torch.Tensor) -> torch.Tensor: ...
def flatten_bn(x: list[torch.Tensor]) -> list[torch.Tensor]: ...
@overload
def flatten_bn(x: list[torch.Tensor]) -> list[torch.Tensor]: ...
def flatten_bn(x: torch.Tensor) -> torch.Tensor: ...
@overload
@@ -973,7 +972,9 @@ def get_draft_quant_config(vllm_config: VllmConfig) -> "QuantizationConfig | Non
The draft model's config if available, None otherwise.
"""
draft_model_config = vllm_config.speculative_config.draft_model_config
speculative_config = vllm_config.speculative_config
assert speculative_config is not None
draft_model_config = speculative_config.draft_model_config
draft_load_config = vllm_config.load_config
return (