Skip to content

[Perf] Optimize the Qwen2.5-Omni Model thinker-to-talker-proj with nn.Linear #825

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hsliuustc0106 merged 1 commit into from
Jan 17, 2026
Merged

[Perf] Optimize the Qwen2.5-Omni Model thinker-to-talker-proj with nn.Linear #825

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .buildkite/pipeline.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ steps:
- "/fsx/hf_cache:/fsx/hf_cache"

- label: "Diffusion Parallelism Test"
timeout_in_minutes: 20
timeout_in_minutes: 25
depends_on: image-build
commands:
- pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
Expand Down
9 changes: 2 additions & 7 deletions vllm_omni/model_executor/models/qwen2_5_omni/qwen2_5_omni_talker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,6 @@
# from vllm.attention import AttentionMetadata # unused import
from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.linear import ColumnParallelLinear
from vllm.model_executor.models.interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
from vllm.model_executor.models.qwen2_5_omni_thinker import (
Qwen2_5OmniThinkerDummyInputsBuilder,
Expand Down Expand Up @@ -68,13 +67,9 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
else:
self.config = config

self.thinker_to_talker_proj = ColumnParallelLinear(
self.thinker_to_talker_proj = nn.Linear(
self.config.embedding_size,
self.config.hidden_size,
bias=True,
gather_output=True,
skip_bias_add=False,
quant_config=quant_config,
)
Comment on lines +70 to 73
Copy link
Copy Markdown

@chatgpt-codex-connector chatgpt-codex-connector Bot Jan 17, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Keep tensor-parallel/quantized linear for projector

Replacing ColumnParallelLinear with nn.Linear drops both tensor-parallel sharding and quant_config handling. In runs with tensor_parallel_size > 1 or quantized checkpoints, this projector will now instantiate full‑precision weights on every rank (no partitioning/quantized kernels), which can inflate memory or prevent quantized weights from loading correctly. Consider keeping the parallel/quantized linear wrapper or providing an equivalent vLLM-aware linear here.

Useful? React with 👍 / 👎.

Copy link
Copy Markdown
Contributor Author

@kechengliu97 kechengliu97 Jan 17, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Meanwhile, Qwen-2.5-Omni-7B is a small model which almost all kinds of NPU/GPU can hold with only one.

self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
Expand Down Expand Up @@ -145,7 +140,7 @@ def forward(
input_ids = None

# projection
inputs_embeds, _ = self.thinker_to_talker_proj(inputs_embeds)
inputs_embeds = self.thinker_to_talker_proj(inputs_embeds)

hidden_states = self.language_model.model(
input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds
Expand Down