ZIBEE/bi_150-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Upgrade to vllm 0.17.0 corex v4.1 overlay

Browse Source
This commit is contained in:
LZiBee
2026-04-29 19:38:22 +08:00
parent 8fac6062e4
commit 938d0854a5
430 changed files with 35969 additions and 14511 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/transformers/base.py
View File

@@ -300,14 +300,26 @@ class Base(
for child_name, child_module in module.named_children():
new_module = child_module
qual_name = maybe_prefix(prefix, child_name)
# Populate Eagle3 attrs
if (
isinstance(module, nn.ModuleList)
and len(module) == self.text_config.num_hidden_layers
):
# Populate Eagle3 attrs
self._target_class = type(child_module)
layer_name = qual_name.removeprefix("model.")
self._layer_names[int(child_name)] = layer_name
# MTP weights should not be loaded into the base model
num_hidden_layers = self.text_config.num_hidden_layers
names = (
"n_predict", # Override from SpeculativeConfig
"num_nextn_predict_layers", # Most models
"mtp_num_hidden_layers", # Qwen 3.5
)
n_predict = getattr_iter(self.text_config, names, 0)
for i in range(num_hidden_layers, num_hidden_layers + n_predict):
mtp_prefix = f"{prefix}.{i}."
if mtp_prefix not in self.ignore_unexpected_prefixes:
self.ignore_unexpected_prefixes.append(mtp_prefix)
# Replace modules as needed
if isinstance(child_module, nn.Linear):
generator = (p for p in tp_plan if re.match(p, qual_name))

26
vllm/model_executor/models/transformers/multimodal.py
View File

@@ -218,7 +218,7 @@ class MultiModalProcessor(BaseMultiModalProcessor[MultiModalProcessingInfo]):
if "mm_token_type_ids" in processed_data
else "token_type_ids"
)
mm_token_type_ids = processed_data.pop(token_type_key)
mm_token_type_ids = processed_data.get(token_type_key)
# We can infer vLLM style placeholder from token type ids, if we split
# it for each input `mm_data`.
@@ -353,6 +353,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
num_image_patches = kwargs.pop("num_image_patches")
kwargs.pop("token_type_ids", None) # used only in `forward`
kwargs.pop("mm_token_type_ids", None) # used only in `model.get_rope_index`
if pixel_values is not None:
# ROCm: Force math SDP backend for vision encoder to avoid accuracy issues
@@ -443,6 +444,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
{
"image_grid_thw",
"video_grid_thw",
"mm_token_type_ids",
"second_per_grid_ts",
"audio_feature_lengths",
"use_audio_in_video",
@@ -451,7 +453,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
if any(
v
for k, v in kwargs.items()
if k not in {"image_grid_thw", "video_grid_thw"}
if k not in {"image_grid_thw", "mm_token_type_ids"}
):
raise NotImplementedError(
"Transformers modeling backend only supports images."
@@ -459,6 +461,7 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
image_grid_thw = kwargs.get("image_grid_thw", [])
video_grid_thw = kwargs.get("video_grid_thw", [])
mm_token_type_ids = kwargs.get("mm_token_type_ids")
image_grid_thw = (torch.stack if image_grid_thw else torch.tensor)(
image_grid_thw
@@ -467,10 +470,29 @@ class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
video_grid_thw
)
# In v4 `get_rope_index` doesn't have wildcard `kwargs`, and
# can't accept arbitrary args, even if its value is `None`
kwargs = {}
if mm_token_type_ids:
if not hasattr(self, "_get_rope_index_accepts_mm_token_type_ids"):
import inspect
sig = inspect.signature(self.model.get_rope_index)
params = sig.parameters
self._get_rope_index_accepts_mm_token_type_ids = (
"mm_token_type_ids" in params
or any(
p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values()
)
)
if self._get_rope_index_accepts_mm_token_type_ids:
kwargs["mm_token_type_ids"] = torch.cat(mm_token_type_ids)
mrope_positions, mrope_position_delta = self.model.get_rope_index(
input_ids=torch.tensor(input_tokens).unsqueeze(0),
image_grid_thw=image_grid_thw,
video_grid_thw=video_grid_thw,
**kwargs,
)
mrope_positions = mrope_positions[:, 0]