[CI] remove benchmark/testing comparison w/ other frameworks#2179
Conversation
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
|
@wtomin @hsliuustc0106 PTAL thanks |
There was a problem hiding this comment.
Pull request overview
Removes SGLang diffusion benchmark/testing comparison from the CI performance benchmarking flow, leaving vLLM-Omni as the only supported backend for these diffusion perf runs.
Changes:
- Deleted the SGLang Qwen-Image diffusion perf test configuration JSON.
- Removed the SGLang server runner + Cache-DiT version checks from the diffusion benchmark CI runner script.
- Dropped SGLang backend support from the diffusion benchmark serving CLI and backend request mappings.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| tests/dfx/perf/tests/test_qwen_image_sglang_diffusion.json | Removes the SGLang-specific diffusion perf test definitions. |
| tests/dfx/perf/scripts/run_diffusion_benchmark.py | Removes SGLang server support and enforces vLLM-Omni-only configs. |
| benchmarks/diffusion/diffusion_benchmark_serving.py | Removes sglang from --backend CLI choices and updates attribution comment. |
| benchmarks/diffusion/backends.py | Removes SGLang request implementations and mapping entries. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| """ | ||
| Performance benchmark CI runner for diffusion models. | ||
|
|
||
| Supports two server backends: | ||
| Supports vLLM-Omni server backend: | ||
| - vllm-omni (default): starts DiffusionServer via vllm_omni.entrypoints.cli.main, | ||
| benchmarks with diffusion_benchmark_serving.py --backend vllm-omni | ||
| - sglang: starts SglangServer via `sglang serve`, | ||
| benchmarks with diffusion_benchmark_serving.py --backend sglang | ||
|
|
||
| A config JSON file is REQUIRED via --config-file: | ||
| pytest run_diffusion_benchmark.py --config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json | ||
| pytest run_diffusion_benchmark.py --config-file tests/dfx/perf/tests/test_qwen_image_sglang_diffusion.json | ||
|
|
||
| JSON config entries are distinguished by a "server_type" field ("vllm-omni" or "sglang"). | ||
| sglang entries support two additional fields under server_params: | ||
| - "env": dict of extra environment variables (e.g. SGLANG_CACHE_DIT_ENABLED) | ||
| - "cache_dit_config": dict written to a temp YAML and passed as | ||
| --cache-dit-config to sglang serve (requires cache-dit == 1.3.0) | ||
| JSON config entries use a "server_type" field, and this runner executes | ||
| the vllm-omni path. |
There was a problem hiding this comment.
The PR description is still the unfilled template (missing Purpose/Test Plan/Test Result), which doesn’t match the scope of this change (CI benchmark backend removal). Please update the PR description to include: (1) why SGLang benchmarking is being removed, (2) what CI/jobs are affected, and (3) what you ran (or why no tests are needed) and the results.
| seen.add(test_name) | ||
| server_type = cfg.get("server_type", "vllm-omni") | ||
| if cfg.get("server_type", "vllm-omni") != "vllm-omni": | ||
| raise ValueError(f"Unsupported server_type in config: {cfg.get('server_type')}") |
There was a problem hiding this comment.
This error message is a bit terse for CI/debugging. Consider including the allowed value(s) and the failing test name to make failures actionable (e.g., mention that only "vllm-omni" is supported and include cfg.get("test_name")).
| raise ValueError(f"Unsupported server_type in config: {cfg.get('server_type')}") | |
| raise ValueError( | |
| f"Unsupported server_type {cfg.get('server_type')!r} in config for test " | |
| f"{cfg.get('test_name', '<unknown>')!r}. Only 'vllm-omni' is supported." | |
| ) |
There was a problem hiding this comment.
The suggestion makes it hard for future additions and flexibility
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Based on occassional fails such as https://buildkite.com/vllm/vllm-omni/builds/5232/steps/canvas Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Based on https://buildkite.com/vllm/vllm-omni/builds/5318/steps/canvas?sid=019d2cf6-1ec5-4d1c-a30f-e1fa841a1162&tab=output Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
https://buildkite.com/vllm/vllm-omni/builds/5326/steps/canvas?sid=019d2d14-786e-47e0-a10a-274764d38354&tab=output Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Updates: remove mixed resolution and save server params into output json file
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
…oject#2179) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
…oject#2179) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
…oject#2179) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: Didan Deng <33117903+wtomin@users.noreply.github.com>
Signed-off-by: Huang, Zeyu 11222265+fhfuih@users.noreply.github.com
PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
Purpose
Test Plan
Test Result
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model. Please runmkdocs serveto sync the documentation editions to./docs.BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)