Fix TurboMind vision workspace overflows

[irexyc]

Summary

This PR prevents InternViT and QwenViT from exceeding the language model's shared symmetric buffer.

Changes

InternViT

• Calculate how many image tiles fit in symm_buf.
• Split oversized vision batches into tile-based micro-batches.
• Write each chunk directly into the final image embedding tensor.
• Fail clearly if the buffer cannot hold one tile.

QwenViT

• Check hidden-state and merger workspace sizes before ViT execution.
• Continue using symm_buf when capacity is sufficient.
• With NCCL, fall back per stage to regular device buffers when necessary.
• With CUDA IPC/native, fail early with detailed capacity information.

The QwenViT change covers Qwen2-VL, Qwen2.5-VL, and Qwen3.5/3.6/3.8 VLMs.

[Copilot]

Pull request overview

This PR addresses TurboMind vision-model workspace overruns by proactively planning symmetric-buffer usage and introducing chunking logic when inputs exceed the language model’s shared symmetric buffer capacity.

Changes:

• QwenViT: pre-compute hidden/merger workspace requirements vs. symm_buf, enforce early failure for non-NCCL communicators when capacity is insufficient, and selectively fall back to regular device buffers where allowed.
• InternViT: compute per-tile workspace requirements and split oversized batches into tile-based micro-batches, writing each chunk directly into the final image_embeds tensor.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/turbomind/models/qwenvit/qwenvit.cc	Adds symmetric-buffer workspace planning and communicator-specific fallback/early-fail behavior.
src/turbomind/models/internvit/internvit.cc	Adds tile micro-batching to keep InternViT intermediate workspaces within symmetric-buffer limits.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.