Skip to content

Support InternS2 Preview#4575

Merged
lvhan028 merged 11 commits into
InternLM:mainfrom
CUHKSZzxy:interns2preview
May 9, 2026
Merged

Support InternS2 Preview#4575
lvhan028 merged 11 commits into
InternLM:mainfrom
CUHKSZzxy:interns2preview

Conversation

@CUHKSZzxy
Copy link
Copy Markdown
Collaborator

No description provided.

Copilot AI review requested due to automatic review settings May 9, 2026 03:14
Copy link
Copy Markdown
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

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

Pull request overview

This PR adds initial support for the InternS2 Preview model family in LMDeploy’s VLM + PyTorch backends, including time-series multimodal inputs and a few related configuration/module-map updates.

Changes:

  • Register InternS2 Preview architectures in VL/PyTorch model registries and module maps.
  • Extend Qwen3.5(-MoE) codepaths to accept and route time-series multimodal inputs end-to-end.
  • Plumb hf_overrides into spec-decode draft model config construction.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 7 comments.

Show a summary per file
File Description
lmdeploy/vl/model/qwen3_5.py Adds InternS2Preview arch + time-series token preprocessing hooks.
lmdeploy/pytorch/models/qwen3_vl.py Adds dtype-aware input processor and TIME_SERIES multimodal packing.
lmdeploy/pytorch/models/qwen3_5.py Adds TIME_SERIES args to forward + generation input preparation.
lmdeploy/pytorch/models/qwen3_5_mtp.py Adds InternS2Preview-specific MTP weight name remapping logic.
lmdeploy/pytorch/models/qwen3_5_moe.py Builds time-series encoder when ts_config exists; adjusts weight loading matching and buffer loading.
lmdeploy/pytorch/models/module_map.py Maps InternS2Preview architectures to existing Qwen3.5 MoE implementation.
lmdeploy/pytorch/models/interns1_pro_time_series.py Removes unused padding_idx.
lmdeploy/pytorch/messages.py Skips mRoPE position-id handling for time-series modality.
lmdeploy/pytorch/engine/config_builder.py Passes hf_overrides into SpecDecodeConfig.from_config.
lmdeploy/pytorch/configurations/qwen3_5.py Recognizes intern_s2_preview model_type in config builder.
lmdeploy/pytorch/config.py Adds hf_overrides plumb-through for spec-decode draft ModelConfig building.
lmdeploy/archs.py Adds InternS2Preview arch to VLM detection + turbomind-unsupported list.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread lmdeploy/vl/model/qwen3_5.py Outdated
Comment thread lmdeploy/vl/model/qwen3_5.py Outdated
Comment thread lmdeploy/pytorch/models/qwen3_vl.py
Comment thread lmdeploy/pytorch/models/qwen3_5.py
Comment thread lmdeploy/pytorch/models/qwen3_5.py Outdated
Comment thread lmdeploy/pytorch/models/qwen3_5_moe.py
Comment thread lmdeploy/archs.py Outdated
@lvhan028 lvhan028 requested a review from RunningLeon May 9, 2026 04:25
@lvhan028 lvhan028 added the enhancement New feature or request label May 9, 2026
Comment thread lmdeploy/pytorch/models/qwen3_5_mtp.py
Copy link
Copy Markdown
Collaborator

@RunningLeon RunningLeon left a comment

Choose a reason for hiding this comment

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

LGTM

@lvhan028 lvhan028 merged commit df9b428 into InternLM:main May 9, 2026
5 checks passed
@CUHKSZzxy CUHKSZzxy deleted the interns2preview branch May 9, 2026 07:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants