[danielhanchen]

Three recent AMD fixes for Qwen-Image-2.1 turn on automatically only on Strix Halo (gfx1151), because that was the only GPU they had been tested on. @LeoBorcherding has since tested all three on a discrete Radeon AI PRO R9700 (RDNA 4, gfx1201). Based on his results and upstream evidence, this PR enables two of them on more Radeon GPUs.

Fix	Before	After	Measured on	Added without a measurement
MIOpen search cutoff (#12359): avoids a multi-minute first VAE decode	gfx1151	RDNA 3, 3.5 and 4	gfx1151, gfx1201	Other RDNA 3, 3.5 and 4
Attention query chunking (#12335): speeds up large generations	gfx1151	gfx1151, gfx1200, gfx1201	gfx1151, gfx1201	gfx1200
Vision eager attention (#12360): fixes blank reference-image results	gfx1151	Unchanged	gfx1151, gfx1201 (no bug)	None

Each section gives the change, Leo's R9700 measurement, the reasoning for anything enabled beyond that measurement, and what stays excluded. His runs used Windows 11, PyTorch 2.11 / ROCm 7.13, and main at 364a42f0, with the existing environment overrides forcing each path on and off.

MIOpen search cutoff (#12359)

Change. MIOPEN_SEARCH_CUTOFF=1 now defaults on when every visible GPU is gfx110*, gfx115* or gfx120*, instead of only gfx1151. Explicit environment values still win, including 0.

Measured. From the #12359 test, with empty MIOpen caches and one clean run per side:

First 1024² generation	Unset	=1
VAE decode	400.0 s	33.5 s
Full render	411.0 s	45.9 s

Images differ by at most 2/255, consistent with the variation between stock runs.

Beyond the measurement. The rest of RDNA 3, 3.5 and 4 is included because the cause is not specific to one GPU. ROCm 7.13 ships no system find databases for any of them, so every first convolution search starts cold. AMD itself sets the cutoff for its cold diffusion benchmarks on Navi31, Navi48 and Strix Halo.

Excluded. Hosts with any other architecture keep MIOpen's default:

• CDNA ships tuned databases. A related naive-solver skip, MIOPEN_NAIVE_DISABLE_IF_ALT, cost 18.8% throughput on MI355X.
• gfx1250 is an Instinct part, which is why the prefixes are not a bare gfx12.
• RDNA 2 ships a database and has no measurement.

The cutoff can discard the only applicable solver and fail a convolution. This risk already applies to the gfx1151 default, and MIOPEN_SEARCH_CUTOFF=0 opts out.

Attention query chunking (#12335)

Change. Automatic 512-row query chunking now covers gfx1200 and gfx1201 as well as gfx1151. The shape and backend guards and the UNSLOTH_QWEN_IMAGE_ROCM_QUERY_CHUNKING overrides are unchanged. The arch is now read with _props_gfx_arch, as the MIOpen and vision gates already do.

Measured. From the #12335 test, four runs per side in alternating order:

1536², 4 steps, transformer forward	Off, median (range)	On, median (range)
Warm	13.41 s (13.19 to 13.49)	11.64 s (11.37 to 12.26)
Cold	19.36 s (18.91 to 19.76)	17.63 s (17.22 to 17.71)
Peak allocation	14.90 GiB	14.90 GiB

Outputs differ by at most 17/255, against 2/255 between stock runs, so chunking changes the numerics slightly.

Beyond the measurement. gfx1200 is included because AOTriton 0.11.2b gives it gfx1201's tuning database and kernel pack, so it runs the same attention kernels.

Excluded. gfx1100, the other RDNA 3 and 3.5 GPUs, and CDNA stay opt-in until they are measured.

Vision eager attention (#12360)

Unchanged. In the #12360 test, native vision SDPA stayed finite on the R9700 with a 1024² reference, with the same speed and memory as eager attention.

Validation

All 237 tests across the MIOpen cutoff, query chunking, diffusion speed and vision attention suites pass. Gate coverage includes mixed hosts, excluded architectures and explicit overrides.

This commit was not run on AMD hardware. gfx1200, discrete RDNA 3 and Linux RDNA 4 remain untested.

---

Mirror of unslothai/unsloth#12471 by @oobabooga, opened for review only.
Do not merge here. Fixes agreed on this mirror must be applied to the source PR.
Source description digest at mirror time: 3e1040750f5b

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-02T07:33:42.679748Z	23e7614	Manual request
🔒 Security Review	✅ Completed 2026-10-02T07:32:56.398853Z	23e7614	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[danielhanchen]

@codex review

[danielhanchen]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

Reviewed commit: 23e7614145

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".