[sergereview]

Original CI failure

• Failure group: 3 integration tests for model deepseek_vl_hybridfailing withother (other (3))
• tests/models/deepseek_vl_hybrid/test_modeling_deepseek_vl_hybrid.py::DeepseekVLHybridIntegrationTest::test_model_text_generation [multi-gpu] (other, seen 21/21)

CI traceback — tests/models/deepseek_vl_hybrid/test_modeling_deepseek_vl_hybrid.py::DeepseekVLHybridIntegrationTest::test_model_text_generation

(line 67)  RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1!

Where to watch it:

• test_model_text_generation — Test dashboard

The three DeepseekVLHybridIntegrationTest text-generation tests crash on multi-gpu with Expected all tensors to be on the same device (cuda:0/cuda:1) and on single-gpu with You can't move a model that has some modules offloaded to cpu or disk / OOM.

Root cause: each test loads the 7B model with device_map="auto" then immediately calls model.to(torch_device). The device map installs accelerate hooks that keep moving submodule inputs to their mapped device; .to() moves weights but does not remove the hooks, so tensors keep crossing devices.

Fix: drop the redundant model.to(torch_device) (inputs already go to model.device), mirroring the merged, GPU-verified, maintainer-approved sibling fix for deepseek_vl (#48536). Applied to all three integration tests plus test_model_text_generation_batched (byte-identical pattern).

Relates to #49275

---

✅ Verified on GPU

serge ran the targeted @slow test(s) on a GPU runner and opened this PR only after they passed with this patch. Each targeted test was run 5× on both the pre-patch and patched trees to rule out flakiness — the result held on every run.

• Passes with this patch: https://github.com/huggingface/transformers/actions/runs/37076887007
• Failed before the fix (base commit): https://github.com/huggingface/transformers/actions/runs/37072362339

Possibly related

Existing issues/PRs mentioning test_model_text_generation (keyword match — not verified to share a root cause):

• #48171 — Fix stale expected values in integration tests (cuda sm_86 / Aug04 regressions) (PR, closed, updated 2026-08-21)
• #44739 — fix series of failed test case for janus model (PR, closed, updated 2026-04-01)
• #44179 — Models with incorrect tokenizer_class in tokenization_config.json tha… (PR, closed, updated 2026-02-23)
• #43187 — 🚨Fix floating-point precision in JanusImageProcessor resize (PR, closed, updated 2026-02-02)
• #41977 — Add Phi3.5 Vision Model (PR, open, updated 2025-12-18)

---

This change was produced automatically by serge from a CI failure report. The patch was generated by an LLM and applied by serge; review before merging.

serge v0.1.0 · model: Qwen/Qwen3.8-27B:novita · 22 LLM turns · 26 tool calls · 1172.5s · 797097 in / 76194 out tokens

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_vl_hybrid

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 37077936827:1
Result: success | Jobs: 5 | Tests: 354 | Failures: 0 | Duration: 38s

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[sergereview]

Closing as superseded by #49335, which targets the same test(s):

• tests/models/deepseek_vl_hybrid/test_modeling_deepseek_vl_hybrid.py::DeepseekVLHybridIntegrationTest::test_model_text_generation

Closed by serge at the request of the nightly CI triage.