[BoarQing]

Summary

The Windows RMT Build fails at "Restore oga-artifacts" on main and every PR. This syncs OGA_PR_PATCHES in windows-rmt.yml with windows-deps.yml.

Related issue or design

None.

Why

windows-rmt.yml restores the OGA artifacts with fail-on-cache-miss: true, keyed by OGA_PR_PATCHES. #1335 changed the list in windows-deps.yml, which builds that cache, to 2685 2691 2705 2712 2792 2801. windows-rmt.yml kept 2685 2691 2705 2712 2763, so it asks for windows-oga-0.17.0-pr2685 2691 2705 2712 2763-ort1.31.0-ortpr, which does not exist. The only Windows OGA cache entry is windows-oga-0.17.0-pr2685 2691 2705 2712 2792 2801-ort1.31.0-ortpr.

What

• windows-rmt.yml: OGA_PR_PATCHES now matches windows-deps.yml, with a comment that the two must agree.

ONNXRUNTIME_VERSION, ONNXRUNTIME_PR_PATCHES and AMDGPU_EP_REF already match.

Test plan

• pre-commit run --files .github/workflows/windows-rmt.yml: all hooks pass.
• gh cache list --key windows-oga shows the entry the new key resolves to.
• This PR's own Windows RMT Build run is the end-to-end check.

Checklist

• The change is focused, or links a design/series explaining its scope.
• Relevant tests were added or updated and the results are documented.
• User-facing or design documentation was updated when needed.
• Substantial AI assistance is disclosed, and I reviewed and understand the result.

AI assistance: the diagnosis and change were made with Cursor (Claude).

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.

[github-actions]

L2 Accuracy Results (EP vs CPU)

Model	Combined L2	Total Elems	Skipped NaN/Inf
conv_test_hybrid	1.12256E-07	64	0
GroupQueryAttention_seq256	59.2561	2621440	0
MatMulNBits_o_seq128	259.905	368640	0
QMoE_seq128	24.2583	368640	0

Threshold: 0.01 | Run: 8246 - Commit: b95ef01

[github-actions]

Binary sizes

Baseline: run 41196920350 - d4f6b4a (dll-sizes)

File	This (MB)	Main (MB)	Delta (MB)
hipgpu.dll	59.43	59.43	0.00
custom_kernels_gfx11-7-generic.dll	5.71	5.71	0.00
custom_kernels_gfx11-generic.dll	5.93	5.93	0.00

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	8.98	7.19	247	3	1160
GroupQueryAttention_seq128	10954.60	2.52751	0	5	167
matmul_down_seq128	739.94	3.35	117	3	268

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	9.00	48.16	250	3	15506

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	9.01	9.80	256	3	15545

OGA Benchmark Results

Model	Warmup	Reps	Prompt Len	Gen Tokens	TTFT (ms)	TPS	Peak Mem (GB)	GPU Mem (GB)
gpt-oss-20b-webgpu-int4-rtn-block-32	1	5	128	128	89.4	83.1	1.25	12.34
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	117.2	49.7	1.15	5.40

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	116	50.1

VLM Benchmark Results

Model	TTFT (ms)	TPS	Output
gemma-4-E4B-it-fp16-int4-kquant-prune-logits	3845.1	7.8	This is a high-quality, outdoor photograph featuring a beautiful, fluffy white dog, likely a West Highland White Terrier or a si
Qwen3.8-27B-fp16-ve-fp16-int4-k_quant-gs128-text-dml-onnx	4876.9	14.7	The user wants a detailed description of the image. Let me carefully observe all elements before writing. Subject: A white do

TTFT / TPS vs main

Baseline: run 41196920350 - d4f6b4a | Gate: TTFT +20% / TPS -20%

Suite	Model	TTFT (ms)	Main	Delta	TPS	Main	Delta
oga	gpt-oss-20b-webgpu-int4-rtn-block-32	89.4	87.3	+2.4%	83.1	83.2	-0.1%
oga	Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	117.2	117.2	0.0%	49.7	50.0	-0.6%
vlm	gemma-4-E4B-it-fp16-int4-kquant-prune-logits	3845.1	3922.7	-2.0%	7.8	7.3	+6.3%
vlm	Qwen3.8-27B-fp16-ve-fp16-int4-k_quant-gs128-text-dml-onnx	4876.9	4814.2	+1.3%	14.7	14.6	+0.1%

Run: 8246 - Commit: b95ef01