[Mypy] Fix mypy typing for Ultravox and Unlimited-OCR models (#58239)

Signed-off-by: Ashraf Bhuiyan <mbhuiyan@redhat.com>
This commit is contained in:
Ashraf Bhuiyan
2026-09-27 21:19:04 -04:00
committed by GitHub
parent 39d49539de
commit b9e2903b89
6 changed files with 36 additions and 27 deletions
-1
View File
@@ -113,7 +113,6 @@ SEPARATE_GROUPS = [
EXCLUDE = [
r"vllm/model_executor/models/[kK]",
r"vllm/model_executor/models/[uU]",
r"vllm/model_executor/models/[vV]",
r"vllm/model_executor/models/[wW]",
r"vllm/model_executor/models/[zZ]",
@@ -1,10 +1,9 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable
from collections.abc import Callable
from functools import wraps
from types import FunctionType
from typing import TYPE_CHECKING
from typing import TYPE_CHECKING, Concatenate, ParamSpec, TypeVar
import torch
@@ -20,13 +19,18 @@ if TYPE_CHECKING:
__all__ = ["set_torchao_reload_attrs", "support_quantized_model_reload_from_hp_weights"]
_P = ParamSpec("_P")
_R = TypeVar("_R")
def set_torchao_reload_attrs(model: torch.nn.Module, model_config: ModelConfig):
model._do_torchao_reload = True
model._model_config = model_config
def support_quantized_model_reload_from_hp_weights(original_load_weights: FunctionType):
def support_quantized_model_reload_from_hp_weights(
original_load_weights: Callable[Concatenate["AutoWeightsLoader", _P], _R],
) -> Callable[Concatenate["AutoWeightsLoader", _P], _R]:
"""Decorator for `load_weights` method for AutoWeightsLoader.load_weights to support
reloading high precision (bfloat16/float16/float32) weight for an already quantized
model, this involves restoring the weights to a high precision weights and
@@ -39,17 +43,16 @@ def support_quantized_model_reload_from_hp_weights(original_load_weights: Functi
@wraps(original_load_weights)
def patched_model_load_weights(
self: "AutoWeightsLoader",
weights: Iterable[tuple[str, torch.Tensor]],
*args,
**kwargs,
):
*args: _P.args,
**kwargs: _P.kwargs,
) -> _R:
model = self.module
if not getattr(model, "_do_torchao_reload", False):
return original_load_weights(self, weights, *args, **kwargs)
return original_load_weights(self, *args, **kwargs)
initialize_layerwise_reload(model)
loaded_weights = original_load_weights(self, weights, *args, **kwargs)
loaded_weights = original_load_weights(self, *args, **kwargs)
finalize_layerwise_reload(model, model._model_config)
return loaded_weights
+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)
+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 (