ci: pin OGA to 0.15.1 and the rebased AMDGPU integration PR

[zz002]

Summary

Moves the OGA pin from v0.14.0 + PR #2194 to the 0.15.1 release commit 3abf4a84 plus PR #2376, the same AMDGPU umbrella integration ported to 0.15.1's device-interface API. Both Windows and Linux workflows are updated, along with the quick-start docs.

Related issue or design

Upstream patch source: microsoft/onnxruntime-genai#2376 (draft, supersedes #2194).

Why

Upstream published no v0.15.1 tag and rel-0.15.1 is a branch that keeps moving, so a tag-shaped ref cannot pin this release: the workflows check out the release commit instead.

The OGA cache key previously combined the PR number with a hand-bumped mm<N> token, which had to be remembered whenever a patch PR's branch changed under the same number. A new step resolves each patch PR's head SHA and folds it into the key, so the cache invalidates on its own when a patch PR is updated.

0.15 enables 1DS telemetry by default, which pulls cpp_client_telemetry (plus curl and mbedtls on Linux) into a cold build and would ship Microsoft telemetry inside AMD artifacts, so OGA now builds with --no_telemetry.

What

• OGA_VERSION (only ever used to build a v<version> ref) is replaced by OGA_COMMIT; OGA_PR_PATCHES moves to 2376.
• New Resolve OGA PR head SHAs step (pwsh on Windows, bash on Linux) queries gh api repos/microsoft/onnxruntime-genai/pulls/<n> --jq .head.sha and emits <pr>-<sha12> pairs as a step output consumed by the cache key.
• --no_telemetry added to both build.py invocations and to the documented local build.
• docs/quick_start.md / docs/quick_start_linux.md: clone-and-checkout the pinned commit, apply PR #2376, and drop the stale morphizen_ep -> amdgpu rename rationale from the git am note.

Test plan

• Locally built OGA at rel-0.15.1 + the #2376 branch against ORT 1.27 with --no_telemetry (Ninja, Release): onnxruntime-genai.dll and model_benchmark.exe produced.
• model_benchmark on Llama-3.1-8B (AWQ int4, "AMDGPU": {"profile": "hip"}, session.disable_cpu_ep_fallback=1) with that build: exit 0, TTFT 252 ms and 30.2 tok/s against 240 ms and 30.9 tok/s for the current CI package (v0.14.0 + #2194).
• Cold-cache CI: Resolve OGA PR head SHAs produces a non-empty SHA that appears in the cache key, Apply OGA PR patches applies #2376 cleanly, Build OGA produces model_benchmark.exe + onnxruntime-genai.dll + the 0.15.1 wheel.
• gpu-test Run OGA benchmark and Run OGA wheel smoke stay green.

Notes for reviewers

PR #2376 is a draft opened as a patch source for this CI; it overlaps with upstream #2165. Once #2165 lands, OGA_PR_PATCHES switches to it and #2376 is closed. The cache key changes on this PR, so the first run is a cold OGA build.

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.

[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	4.8668E-07	64	0
GroupQueryAttention_seq256	25.2366	2621440	0
MatMulNBits_o_seq128	259.906	368640	0
QMoE_seq128	34.9552	368640	0

Threshold: 0.01 | Run: 3668 - Commit: 18be2ca

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.27	366	3	1244
GroupQueryAttention_seq128	4469.93	1.6761	10	6	311
matmul_down_seq128	522.06	2.31	69	3	352

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.54	45.97	365	3	15590

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.53	9.85	363	3	15760

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	172.6	77.6	1.33	13.53
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	344.1	41.2	1.22	6.43

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	196	40.4

Run: 3668 - Commit: 18be2ca