[codex] add qwen3.5-4b inference example

[Fatemanx]

🗂️ PR Category

• ✨ New Feature

📝 Description

Add a forward-only Qwen3.5-4B inference example under example/inference/qwen3.5-4b.

What changed

• Adds infer.py, infer.sh, modeling.py, and a short README for the new example.
• Keeps only the minimal compiled entrypoints: text_prefill, text_decode, image_prefill, image_decode, and all.
• Requires MODEL_PATH from the environment; no local weight path, GPU id, or Cutlass default is baked in.
• Follows the same minimal VAE-style wrapper pattern already used by the repo's inference examples.

Validation

• python -m py_compile example/inference/qwen3.5-4b/infer.py example/inference/qwen3.5-4b/modeling.py
• bash -n example/inference/qwen3.5-4b/infer.sh
• pytest tests/example_inference/test_qwen35_4b_infer.py -q -> 6 passed

[wtr0504]

@Fatemanx Hello! It looks like the CI is failing on this PR. Could you take a look?

[wtr0504]

@Fatemanx Thanks a lot for your contribution! We really appreciate it. We'll consider merging this PR, and let's work together to get it across the finish line. Could you rebase your branch onto main?

[Fatemanx]

@Fatemanx Thanks a lot for your contribution! We really appreciate it. We'll consider merging this PR, and let's work together to get it across the finish line. Could you rebase your branch onto main?

I rebase to main already.

[wtr0504]

The current harness hardcodes batch=1 ((1, seq_len) / (1, 1)). For a compiler example this is too narrow — MagiCompiler needs to see a real batch dimension.

[wtr0504]

infer.sh already wraps nsys. Please run each mode on a real GPU and paste the nsys screenshots into the PR (timeline / CUDA HW view is enough)