[Tracking] Release v0.1.13 validation (DO NOT MERGE)

[sunway513]

Purpose

Tracking-only PR to fire ci:all extended validation against release/v0.1.13.

DO NOT MERGE.

Branch state

Commit	Source	Notes
74dcf38d	this branch	ci(release): cherry-pick #2875 — torch_pin + torch_index_url workflow inputs
e039817a	main 2026-04-23 11:15Z	MLA PS mode support nhead8,2 in MI308 (#2852) — release base

Why this commit as v0.1.13 base

• ATOM benchmark dashboard shows ATOM commit 4f3c57970c (2026-04-23 14:46Z) ran 14/14 accuracy pass against this AITER head.
• ATOM commit 9ea09987bf (2026-04-23 13:05Z) also 14/14 pass against this AITER head.
• AITER's own atom-test CI on e039817a is 3/3 green (DSR1 MI300X + MI355X, gpt-oss-120b MI355X).
• Avoids the 04-24 churn around fmha / sliding-window-mtp / gpt-oss-tuned-config which had cancelled / failed CI.

Validation matrix (target)

• aiter-test (1 GPU, 8 GPU)
• sglang_downstream
• vllm benchmark — no regression vs post1
• atom-test (DSR1, GPT-OSS, etc.)
• triton-test (MI355)
• Manual GSM8K accuracy (5 models)
• vLLM + ATOM plugin OOT
• SGLang + ATOM plugin OOT
• 6-wheel build matrix (manylinux + torch 2.10)

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2913 --add-label <label>

[Copilot]

Pull request overview

Tracking PR to run extended CI validation for release/v0.1.13, including a cherry-pick that makes the release workflow’s manylinux build able to pin/override the PyTorch wheel source for ABI-targeted builds.

Changes:

• Add torch_pin and torch_index_url inputs to workflow_dispatch and workflow_call.
• Wire the new inputs into job environment variables.
• Update the manylinux dependency install step to optionally use the provided index URL and torch version spec.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

TORCH_INDEX_URL is user-provided now, but it’s accepted verbatim as the pip --index-url. To reduce supply-chain risk and accidental misuse, consider validating this input before use (e.g., require https:// and/or restrict to an allowlist like download.pytorch.org unless an explicit “unsafe_allow_custom_index” flag is set).

Suggested change

	# legacy auto-derived behavior.
	if [ -n "${TORCH_INDEX_URL}" ]; then
	TORCH_INDEX="${TORCH_INDEX_URL}"
	else
	TORCH_INDEX="https://download.pytorch.org/whl/rocm${ROCM_NUM}"
	# legacy auto-derived behavior. Custom indexes must use HTTPS and
	# are restricted to download.pytorch.org unless explicitly allowed.
	DEFAULT_TORCH_INDEX="https://download.pytorch.org/whl/rocm${ROCM_NUM}"
	if [ -n "${TORCH_INDEX_URL}" ]; then
	case "${TORCH_INDEX_URL}" in
	https://*)
	;;
	*)
	echo "TORCH_INDEX_URL must start with https://" >&2
	exit 1
	;;
	esac
	case "${TORCH_INDEX_URL}" in
	https://download.pytorch.org/*)
	TORCH_INDEX="${TORCH_INDEX_URL}"
	;;
	*)
	if [ "${UNSAFE_ALLOW_CUSTOM_TORCH_INDEX}" = "true" ]; then
	TORCH_INDEX="${TORCH_INDEX_URL}"
	else
	echo "TORCH_INDEX_URL host is not allowed. Set UNSAFE_ALLOW_CUSTOM_TORCH_INDEX=true to permit a custom HTTPS index." >&2
	exit 1
	fi
	;;
	esac
	else
	TORCH_INDEX="${DEFAULT_TORCH_INDEX}"

[Copilot]

Since TORCH_INDEX can now come from the caller-provided torch_index_url, echoing it verbatim can leak credentials if someone passes a URL containing embedded auth (or other sensitive query params). Consider either rejecting URLs with credentials, or logging a sanitized/redacted form of the URL.

Suggested change

	# Optional torch version pin (e.g. 2.10.0+rocm7.1). Empty = latest.
	if [ -n "${TORCH_PIN}" ]; then
	TORCH_SPEC="torch==${TORCH_PIN}"
	else
	TORCH_SPEC="torch"
	fi
	echo "Torch index: ${TORCH_INDEX}"
	SAFE_TORCH_INDEX="$(TORCH_INDEX="${TORCH_INDEX}" python3 - <<'PY'
	import os
	from urllib.parse import urlsplit, urlunsplit
	url = os.environ["TORCH_INDEX"]
	parts = urlsplit(url)
	if not parts.scheme or not parts.netloc:
	print(url)
	else:
	host = parts.hostname or ""
	if parts.port is not None:
	host = f"{host}:{parts.port}"
	print(urlunsplit((parts.scheme, host, parts.path, "", "")))
	PY
	)"
	# Optional torch version pin (e.g. 2.10.0+rocm7.1). Empty = latest.
	if [ -n "${TORCH_PIN}" ]; then
	TORCH_SPEC="torch==${TORCH_PIN}"
	else
	TORCH_SPEC="torch"
	fi
	echo "Torch index: ${SAFE_TORCH_INDEX}"