[syang-genai]

Motivation: fixes #2865. Qwen3.5 image prompts reach the runner without embedding_spans, so they take the whole-image merge in ModelRunner.run_model: every vision row is scattered into the forward's image tokens. After a prefix-cache hit that ends inside or past the image, the forward holds fewer image tokens than the vision tower returns rows, and merge_multimodal_embeddings fails with a shape mismatch ([784, 2048] into [0, 2048] in the issue's A, B, A repro).

Technical details:

• build_qwen_vl_inputs returns each image's run of image tokens as an embedding_spans entry, plus the media cache_seed. This applies to the qwen3_5 and qwen3_5_moe model types, the same gate as the Qwen3.5 MRoPE path. The runner then takes the span path that already exists: it scatters only the slice of each image a forward contains, and the image prompt no longer needs one atomic prefill.
• qwen3_5_image_spans walks the prompt and claims each image's t*h*w / merge**2 tokens in turn, so adjacent images without separators split correctly. Prompts whose image tokens don't match the grids are rejected when the request is built, instead of failing in the forward.
• Qwen3.8 (qwen4_exp) stays on the whole-image path. Its processor output differs, and I'd rather handle it separately.

Test Plan:

• tests/test_multimodal_inputs.py:
  • one span per image, for both model types;
  • cache_seed tells different images apart;
  • prompts that disagree with the grid are rejected;
  • other Qwen model types keep the whole-image merge.
• tests/test_multimodal_runtime.py:
  • after prefix hits ending before, inside and past the image, the forward scatters only the image rows left in it;
  • an image prompt is chunked and resumes after a prefix hit.

python -m pytest tests/test_multimodal_inputs.py tests/test_multimodal_runtime.py

Test Result:

• On this branch: 69 passed.
• With atom/models/qwen3_5_vl.py reverted to main: 13 of the 14 new tests fail. The one that passes is the check that other model types are unchanged.
• black and ruff are clean apart from two findings that are already on main.

Not done yet: I haven't run an end-to-end check on MI3xx/MI355X. I don't have one, and ATOM has no gfx11 path to test on my machine. If someone could rerun the issue's ATOM_GDN_STATE_MIDSTEP=1 A, B, A repro on this branch, that would close the gap. Opening as a draft until then.

@ZLkanyo009, if you'd rather fold this into #2636, I'll close this one.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 3041 --add-label <label>