[AMD] add dsv4 sglang disagg

[billishyahao]

cc @Duyi-Wang

---

Note

Medium Risk
Touches shared disagg launch paths (server_sglang.sh, models.yaml) for all models, not only DSv4; behavior changes when EP is disabled and MoE auto-sizing is partially commented out.

Overview
Adds dsv4-fp4-mi355x-sglang-disagg to the AMD master benchmark matrix (8k/1k, non-MTP) with sweeps over pure TP8, DEP8 (MoRI KV + MoE a2a), and dp-attention + TP-MoE, plus a new workflow runner dsv4_fp4_mi355x_sglang-disagg.sh and a perf-changelog entry.

The multi-node harness is extended for DSv4 PD: a DeepSeek-V4-Pro block in models.yaml (dsv4 attention backend, mori disagg, prefill disable_cuda_graph) and matching MoRI/kernel env overrides in env.sh; the bench client uses --dsv4 framing instead of chat templates.

server_sglang.sh / models.yaml refactor MoE CLI so ep_flags (mori a2a, deepep, fake dispatch) apply only when EP is on—ep=1 stays TP-MoE even with dp-attention—and prefill can honor per-model disable_cuda_graph, context_length, and optional MORI_NUM_MAX_DISPATCH_TOKENS_PER_RANK_* overrides. submit.sh threads DRY_RUN for previewing composed launch commands on a real allocation.

^{Reviewed by Cursor Bugbot for commit f56f8de. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c22652b. Configure here.}

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27731746053
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27731746053

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27731746053
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27731746053

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27731746053
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27731746053

[functionstackx]

can u also ur ai agent include descriptions + links of some of the bug fix PRs in here like

• [AMD] Support unified_kv_triton for disaggregation sgl-project/sglang#27935
• [AMD] fix moriep quant kernel not implemented issue sgl-project/sglang#27855
• [PD][MoRI] Align hybrid state transfer with per-component schema sgl-project/sglang#26539

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27893589025
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27893589025

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27896968169
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=27896968169

[functionstackx]

hi @billishyahao there seems to be an accuracy issues with TP8+TP8. codex has narrowed it down to conc=4, here is the bug report for when u wake up, please take a look

sgl-project/sglang#28851

https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27896968169/job/82550287079?pr=1818

[functionstackx]

can u remove https://github.com/SemiAnalysisAI/InferenceX/blob/main/benchmarks/multi_node/amd_utils/patches/mori_conn.py too

this is no longer needed now that sgl-project/sglang#26525 is fixed in sgl-project/sglang#26539

InferenceX/benchmarks/multi_node/amd_utils/job.slurm

Lines 73 to 80 in 6a07901

	_MORI_PATCH_FILE="$DI_REPO_DIR/benchmarks/multi_node/amd_utils/patches/mori_conn.py"
	_MORI_PATCH_TARGET="/sgl-workspace/sglang/python/sglang/srt/disaggregation/mori/conn.py"
	if [[ "${MORI_CONN_PATCH:-auto}" != "skip" ]] \
	&& [[ -f "$_MORI_PATCH_FILE" ]] \
	&& [[ "${DOCKER_IMAGE_NAME:-}" == *"v0.5.12.post1"* ]] \
	&& [[ "${EXTRA_DOCKER_MOUNTS:-}" != *"$_MORI_PATCH_TARGET"* ]]; then
	EXTRA_DOCKER_MOUNTS="${EXTRA_DOCKER_MOUNTS:-} -v ${_MORI_PATCH_FILE}:${_MORI_PATCH_TARGET}:ro"
	export EXTRA_DOCKER_MOUNTS