minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] fix: nightly diffusion benchmark passes the retired --warmup flag (#34423)

Browse Source 
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Mick
2026-08-12 10:25:14 +08:00
committed by GitHub
co-authored by Claude Opus 5
parent 30cb848d4b
commit 81c88da1ab
2 changed files with 14 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/ci/utils/diffusion/comparison_configs.json
+11 -11
View File
@@ -13,7 +13,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --dit-layerwise-offload false --tp-size 2", "serve_args": "--warmup-mode server --dit-layerwise-offload false --tp-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -29,7 +29,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --dit-layerwise-offload false --tp-size 2", "serve_args": "--warmup-mode server --dit-layerwise-offload false --tp-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -45,7 +45,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --tp-size 2", "serve_args": "--warmup-mode server --tp-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -62,7 +62,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --tp-size 2", "serve_args": "--warmup-mode server --tp-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -78,7 +78,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --tp-size 2", "serve_args": "--warmup-mode server --tp-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -95,7 +95,7 @@
"num_gpus": 4, "num_gpus": 4,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory", "serve_args": "--warmup-mode server --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
"extra_env": {} "extra_env": {}
} }
} }
@@ -113,7 +113,7 @@
"num_gpus": 1, "num_gpus": 1,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup", "serve_args": "--warmup-mode server",
"extra_env": {} "extra_env": {}
} }
} }
@@ -131,7 +131,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --pipeline-class-name LTX2TwoStagePipeline --cfg-parallel-size 2", "serve_args": "--warmup-mode server --pipeline-class-name LTX2TwoStagePipeline --cfg-parallel-size 2",
"extra_env": {} "extra_env": {}
} }
} }
@@ -147,7 +147,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --tp-size 2 --attention-backend fa", "serve_args": "--warmup-mode server --tp-size 2 --attention-backend fa",
"extra_env": {} "extra_env": {}
} }
} }
@@ -164,7 +164,7 @@
"num_gpus": 2, "num_gpus": 2,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --tp-size 2", "serve_args": "--warmup-mode server --tp-size 2",
"extra_env": {"SGLANG_DISABLE_COSMOS3_GUARDRAILS": "1"} "extra_env": {"SGLANG_DISABLE_COSMOS3_GUARDRAILS": "1"}
} }
} }
@@ -182,7 +182,7 @@
"num_gpus": 4, "num_gpus": 4,
"frameworks": { "frameworks": {
"sglang": { "sglang": {
"serve_args": "--warmup --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory", "serve_args": "--warmup-mode server --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory",
"extra_env": {} "extra_env": {}
} }
} }
scripts/ci/utils/diffusion/run_comparison.py
+3 -3
View File
@@ -802,9 +802,9 @@ def run_single(
wait_for_health(base_url, framework) wait_for_health(base_url, framework)
# No client-side warmup: each framework relies on its own server-side # No client-side warmup: each framework relies on its own server-side
# warmup before traffic. sglang's serve_args pass --warmup, which `serve` # warmup before traffic. sglang's serve_args pass --warmup-mode server,
# resolves to server-based (synthetic) warmup that primes kernels at # which primes kernels with a synthetic request at startup, before the
# startup, before the health check passes. This goes through the internal # health check passes. This goes through the internal
# warmup path that bypasses sampling-param preset validation (e.g. # warmup path that bypasses sampling-param preset validation (e.g.
# Ideogram-4's preset-locked num_inference_steps), so no per-case warmup # Ideogram-4's preset-locked num_inference_steps), so no per-case warmup
# special-casing is needed here. # special-casing is needed here.