[BugFix] Qwen-image performance regressed - Avoid mapping diffusion_batch_size onto scheduler max_num_seqs

[NumberWan]

Fix Qwen-image performance Regression

Fixes

#6435

Summary

#5676 wrote od_config.max_num_seqs = batch_size on the shared diffusion stage init path. batch_size here is diffusion_batch_size (default 1, not bound to --max-num-seqs). Qwen-Image nightly serve uses --step-execution --max-num-seqs 8, so capacity became 1 and HTTP concurrency 8 queued serially.

This PR deletes those two assignments. Scheduler capacity is only --max-num-seqs / stage YAML max_num_seqs. diffusion_batch_size stays a client knob.

Wan request-level batching of width 4 must set --max-num-seqs 4 (or YAML). The default deploy file vllm_omni/deploy/wan2_2_ti2v.yaml now has max_num_seqs: 4. Replica-DP Wan YAML is unchanged (max_num_seqs: 1, one request per replica GPU).

There is no shared max(CLI, batch_size) helper.

Nightly symptom (H100, from #6435)

test_qwen_image_single_device_step_execution_high_concurrency, c=8, n=160:

Metric	Current	Baseline	Change
throughput_qps	0.4799	0.8159	−41.2%
e2e_latency_ms	16309	9656	+68.9%

At c=2/4/8, QPS stayed ~flat (~0.47) while latency scaled nearly linearly (~4.2s → 8.3s → 16.3s). That is serial admission, not a gradual compute regression.

RCA (1× L20X, relative only)

Same case: 512×512, 20 steps, c=8, n=160, --step-execution --max-num-seqs 8.

First slowing commit: 2a315e1a — [Perf] Support request-level batching for Wan2.2 pipelines (#5676). Parent d1e230c9 is still fast. Later days stay slow; no second independent ≥15% drop.

Daily checkpoint (same case):

Day tip	QPS	Mean latency
8/15 (still good)	~0.70	~11.3s
8/16 (first bad day)	~0.53	~14.8s
8/20 (issue tip)	~0.51	~15.4s

A/B (4 runs each):

Variant	Median QPS	Median mean latency	vs parent
Parent d1e230c9	0.723	10.95s	reference
Culprit 2a315e1a	0.502	15.59s	QPS −30.6%, latency +42.4%

CV ~2–3%. Delete only od_config.max_num_seqs = batch_size on the culprit → QPS 0.712 / 11.09s (parent-level). Isolates the overwrite, not Wan pipeline files.

Knob	Meaning	Qwen nightly
--max-num-seqs	scheduler in-flight	8
diffusion_batch_size	client request-batch width (#1593)	default 1

Behavior change

Call sites that relied on diffusion_batch_size=N alone to raise scheduler in-flight width will now stay at CLI/YAML max_num_seqs (often 1). Pass --max-num-seqs N as well.

Test plan

• Unit tests: shared init batch_size=4 leaves max_num_seqs at the CLI/YAML value (1 / 8 / None), not 4.
• Qwen-Image step-execution sweep c=1,2,4,8 (512×512 / 20 steps / n=160). Expect QPS to scale with c, not stay flat at ~0.47.
• Wan request-level batching with --max-num-seqs 4 (YAML or CLI). diffusion_batch_size alone must not raise scheduler capacity.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/stage_runtime.md.

Module owners: @tzhouam @fake0fan

@NumberWan, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[Gaohan123]

LGTM. Thanks