[diffusion] chore: retire released warmup and decoder flags (#33850)

docs/cookbook/diffusion/SANA-WM/SANA-WM.mdx

@@ -207,7 +207,7 @@ Notes on launch behavior:
 - **CPU offload flags are optional.** `--dit-cpu-offload`, `--text-encoder-cpu-offload`, and `--image-encoder-cpu-offload` are available; defaults are auto-adjusted from GPU memory (GPUs under 30 GB get more aggressive offloading).
 - **Multi-GPU realtime.** Prefer explicit sequence parallelism (`--sp-degree` equal to the number of GPUs for a single session). Do not enable CFG parallel for the realtime profile: the default realtime request uses `guidance_scale=1.0`, while CFG parallel requires active cond/uncond branches.
 - **FSDP.** Use `--use-fsdp-inference` only when you specifically need weight sharding for memory. For the low-latency realtime profile, prefer keeping components resident and using SP first.
-- **Warmup.** Server warmup is **automatically skipped** for the realtime pipeline — a synthetic warmup request has no WebSocket session, so the server detects the registered realtime adapter and skips it. No `--warmup` flag is needed.
+- **Warmup.** Server warmup is **automatically skipped** for the realtime pipeline — a synthetic warmup request has no WebSocket session, so the server detects the registered realtime adapter and skips it. No explicit `--warmup-mode` setting is needed.
 
 Once up, the realtime WebSocket endpoint lives at `ws://127.0.0.1:30000/v1/realtime_video/generate` (use the Python client in §7 to connect — plain `curl` does not speak the `ws://` upgrade).
 

docs/docs/sglang-diffusion/cache_dit.mdx

@@ -254,7 +254,7 @@ Then, apply the quantization config from yaml. Please also enable torch.compile
 sglang generate \
   --backend diffusers \
   --model-path Qwen/Qwen-Image \
-  --warmup \
+  --warmup-mode request \
   --cache-dit-config quantize.yaml \
   --enable-torch-compile \
   --dit-cpu-offload false \

docs/docs/sglang-diffusion/disaggregation.mdx

@@ -273,10 +273,6 @@ Set `--disagg-p2p-hostname` to the actual IP on each machine. For multi-machine,
       <td><code>--decoder-sp</code></td>
       <td>Decoder sequence parallelism</td>
     </tr>
-    <tr>
-      <td><code>--decoder-tp</code></td>
-      <td>Deprecated alias for <code>--decoder-sp</code></td>
-    </tr>
   </tbody>
 </table>
 

docs/docs/sglang-diffusion/models_with_ar.mdx

@@ -122,7 +122,7 @@ SGLANG_CACHE_DIT_FN=2 SGLANG_CACHE_DIT_BN=1 SGLANG_CACHE_DIT_WARMUP=4 SGLANG_CAC
 SGLANG_CACHE_DIT_MC=4 SGLANG_CACHE_DIT_TAYLORSEER=true SGLANG_CACHE_DIT_TS_ORDER=2 \
 SGLANG_CACHE_DIT_ENABLED=true sglang generate --model-path /path/to/zai-org/GLM-Image/ \
 --prompt "A curious raccoon" --height 1920 --width 1088 --num-inference-steps 50 --num-gpus 4 \
---sp-degree 4 --srt-encoder-url "http://${HOST}:${PORT}" --warmup
+--sp-degree 4 --srt-encoder-url "http://${HOST}:${PORT}" --warmup-mode request
 ```
 Result:
 ```bash

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/benchmark-and-profile.md

@@ -292,7 +292,7 @@ sglang generate \
   --prompt="At night, while their owner sleeps in a bedroom, three cats march in loudly playing tiny brass instruments, then abruptly file out." \
   --seed=1101 --num-gpus=4 --tp-size=2 --ulysses-degree=2 \
   --performance-mode=speed --enable-torch-compile=false \
-  --save-output --warmup \
+  --save-output --warmup-mode request \
   --perf-dump-path="${BENCH_DIR}/minimax-h3-t2va-baseline.json"
 ```
 
@@ -344,7 +344,7 @@ sglang generate \
   --width=768 --height=512 \
   --num-frames=121 \
   --seed=42 --num-gpus=2 --enable-cfg-parallel \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 `LTX2TwoStagePipeline` is a native path. The spatial upsampler and distilled
@@ -361,7 +361,7 @@ sglang generate \
   --width=768 --height=512 \
   --num-frames=121 \
   --seed=42 --num-gpus=2 --cfg-parallel-size=2 \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 This matches the nightly comparison case `ltx2.3_twostage_ti2v_2gpus`.
@@ -377,7 +377,7 @@ sglang generate \
   --num-frames=121 --fps=24 \
   --num-inference-steps=30 --guidance-scale=3.0 \
   --seed=1234 --num-gpus=2 \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 Use this when you want the native `LTX2Pipeline` baseline for `LTX-2.3` at the
@@ -395,7 +395,7 @@ sglang generate \
   --num-frames=121 --fps=24 \
   --num-inference-steps=30 --guidance-scale=3.0 \
   --seed=1234 --num-gpus=2 \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 This matches the skill-only `ltx23-two-stage` preset. Use it as a
@@ -413,7 +413,7 @@ sglang generate \
   --num-inference-steps=40 --guidance-scale=4.0 \
   --num-gpus=2 --enable-cfg-parallel --ulysses-degree=1 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 ### Manual command example: FireRed Image Edit
@@ -428,7 +428,7 @@ sglang generate \
   --num-inference-steps=40 --guidance-scale=4.0 \
   --num-gpus=2 --enable-cfg-parallel --ulysses-degree=1 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 Use `FireRedTeam/FireRed-Image-Edit-1.0` in the same command when comparing the
@@ -453,7 +453,7 @@ sglang generate \
   --config="${CONFIG_DIR}/hunyuan3d-shape.json" \
   --num-inference-steps=50 --guidance-scale=5.0 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --save-output --enable-torch-compile --warmup
+  --save-output --enable-torch-compile --warmup-mode request
 ```
 
 For Hunyuan3D, compare the denoise stage separately from mesh export and paint
@@ -473,7 +473,7 @@ sglang generate \
   --seed=42 --save-output \
   --num-gpus=4 --enable-cfg-parallel --ulysses-degree=2 \
   --text-encoder-cpu-offload --pin-cpu-memory \
-  --warmup --enable-torch-compile
+  --warmup-mode request --enable-torch-compile
 ```
 
 `Wan2.2-I2V-A14B` uses the 720p max-area config by default, and explicit
@@ -485,7 +485,7 @@ reference-image aspect ratio.
 For every benchmark run, write a perf dump JSON:
 
 ```bash
-sglang generate ... --warmup --perf-dump-path "${BENCH_DIR}/<result>.json"
+sglang generate ... --warmup-mode request --perf-dump-path "${BENCH_DIR}/<result>.json"
 ```
 
 Before/after comparison:
@@ -542,7 +542,7 @@ sglang generate \
   --model-path=black-forest-labs/FLUX.1-dev \
   --prompt="A futuristic cyberpunk city at night" \
   --width=1024 --height=1024 --num-inference-steps=50 \
-  --seed=42 --enable-torch-compile --warmup \
+  --seed=42 --enable-torch-compile --warmup-mode request \
   --profile
 ```
 

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-benchmark-profile/scripts/bench_diffusion_denoise.py

@@ -665,7 +665,7 @@ def _expected_nightly_cli_args(case: dict) -> dict[str, str]:
     serve_args = shlex.split(case["frameworks"]["sglang"].get("serve_args", ""))
     parsed_serve_args = _parse_cli_args(serve_args)
     for flag, value in parsed_serve_args.items():
-        if flag in {"enable-torch-compile", "warmup"}:
+        if flag in {"enable-torch-compile", "warmup-mode"}:
             continue
         expected[flag] = _normalize_cli_value(value)
 
@@ -805,7 +805,7 @@ def build_sglang_cmd(
     if save_output:
         cmd.append("--save-output")
     if warmup:
-        cmd.append("--warmup")
+        cmd.extend(["--warmup-mode", "request"])
     if torch_compile and not cfg.get("force_eager", False):
         cmd.append("--enable-torch-compile")
     if perf_dump_path:

python/sglang/multimodal_gen/.claude/skills/sglang-diffusion-performance/SKILL.md

@@ -32,8 +32,8 @@ These options are intended to preserve output quality. In practice, some paths (
 
 | Option | CLI Flag / Env Var | What It Does | Speedup | Limitations / Notes |
 |---|---|---|---|---|
-| **torch.compile** | `--enable-torch-compile` | Applies `torch.compile` to the DiT forward pass, fusing ops and reducing kernel launch overhead. | ~1.2–1.5x on denoising | First request is slow (compilation). May cause minor precision drifts due to [PyTorch issue #145213](https://github.com/pytorch/pytorch/issues/145213). Pair with `--warmup` for best results. |
-| **Warmup** | `--warmup` | Runs dummy forward passes to warm up CUDA caches, JIT, and `torch.compile`. Eliminates cold-start penalty. | Removes first-request latency spike | Adds startup time. Without `--warmup-resolutions`, warmup happens on first request. |
+| **torch.compile** | `--enable-torch-compile` | Applies `torch.compile` to the DiT forward pass, fusing ops and reducing kernel launch overhead. | ~1.2–1.5x on denoising | First request is slow (compilation). May cause minor precision drifts due to [PyTorch issue #145213](https://github.com/pytorch/pytorch/issues/145213). Pair with `--warmup-mode request` for best results. |
+| **Warmup** | `--warmup-mode request` | Runs dummy forward passes to warm up CUDA caches, JIT, and `torch.compile`. Eliminates cold-start penalty. | Removes first-request latency spike | Adds startup time. Without `--warmup-resolutions`, warmup happens on first request. |
 | **Warmup Resolutions** | `--warmup-resolutions 256x256 720x720` | Pre-compiles and warms up specific resolutions at server startup (instead of lazily on first request). | Faster first request per resolution | Each resolution adds to startup time. Serving mode only; useful when you know your target resolutions in advance. |
 | **Multi-GPU (SP)** | `--num-gpus N --ulysses-degree N` | Sequence parallelism across GPUs. Shards sequence tokens (not frames) to minimize padding. | Near-linear scaling with N GPUs | Requires NCCL; inter-GPU bandwidth matters. `ulysses_degree * ring_degree = sp_degree`. For Wan2.2 video, start by benchmarking pure Ulysses before assuming a mixed Ulysses/Ring layout is fastest. |
 | **CFG Parallel** | `--enable-cfg-parallel` | Runs conditional and unconditional CFG branches in parallel across GPUs. For CFG models on multi-GPU, benchmark this against pure Ulysses on your topology instead of assuming one always wins. | Often faster than pure SP for CFG models | Requires `num_gpus >= 2`. Halves the Ulysses group size (e.g. 8 GPU → two 4-GPU groups). Only for models that use CFG. Nightly coverage configs may intentionally use smaller Ulysses groups to keep ring behavior exercised; that does not automatically make them the lowest-latency choice. |
@@ -127,7 +127,7 @@ Current H3 restrictions:
 ```bash
 sglang generate --model-path Wan-AI/Wan2.2-T2V-A14B-Diffusers \
   --num-gpus 8 --enable-cfg-parallel --ulysses-degree 4 \
-  --enable-torch-compile --warmup \
+  --enable-torch-compile --warmup-mode request \
   --text-encoder-cpu-offload true \
   --prompt "..." --save-output
 ```
@@ -149,7 +149,7 @@ sglang generate --model-path Lightricks/LTX-2 \
   --width 768 --height 512 \
   --num-frames 121 \
   --seed 42 --num-gpus 2 --enable-cfg-parallel \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 Note: LTX-2 is a current-source benchmark preset rather than a nightly
@@ -168,7 +168,7 @@ sglang generate --model-path Lightricks/LTX-2.3 \
   --width 768 --height 512 \
   --num-frames 121 \
   --seed 42 --num-gpus 2 --cfg-parallel-size 2 \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 Note: this matches the nightly comparison case `ltx2.3_twostage_ti2v_2gpus`. The nightly config omits explicit steps and guidance, so this command omits them too and uses runtime defaults. Download `${ASSET_DIR}/cat.png` with the benchmark/profile skill before running it.
@@ -183,7 +183,7 @@ sglang generate --model-path Lightricks/LTX-2.3 \
   --num-frames 121 --fps 24 \
   --num-inference-steps 30 --guidance-scale 3.0 \
   --seed 1234 --num-gpus 2 \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 Note: use this as the native `LTX2Pipeline` baseline for `LTX-2.3`. It keeps the validated one-stage resolution and explicit `LTX-2.3` sampling defaults, and matches the `ltx23-one-stage` benchmark preset in `sglang-diffusion-benchmark-profile`.
@@ -199,7 +199,7 @@ sglang generate --model-path Lightricks/LTX-2.3 \
   --num-frames 121 --fps 24 \
   --num-inference-steps 30 --guidance-scale 3.0 \
   --seed 1234 --num-gpus 2 \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 Note: this is a high-resolution stress target for the native `LTX-2.3` two-stage path. It matches the skill-only `ltx23-two-stage` benchmark preset, not a nightly comparison case.
@@ -208,7 +208,7 @@ Note: this is a high-resolution stress target for the native `LTX-2.3` two-stage
 
 ```bash
 sglang generate --model-path <IMAGE_MODEL> \
-  --enable-torch-compile --warmup \
+  --enable-torch-compile --warmup-mode request \
   --dit-layerwise-offload false \
   --dit-cpu-offload false \
   --prompt "..." --save-output
@@ -227,7 +227,7 @@ sglang generate --backend=sglang \
   --num-inference-steps 40 --guidance-scale 4.0 \
   --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 ```bash
@@ -239,7 +239,7 @@ sglang generate --backend=sglang \
   --num-inference-steps 40 --guidance-scale 4.0 \
   --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 Use `FireRedTeam/FireRed-Image-Edit-1.0` in the same command when comparing
@@ -264,7 +264,7 @@ sglang generate --backend=sglang \
   --config "${CONFIG_DIR}/hunyuan3d-shape.json" \
   --num-inference-steps 50 --guidance-scale 5.0 \
   --dit-layerwise-offload false --dit-cpu-offload false \
-  --enable-torch-compile --warmup --save-output
+  --enable-torch-compile --warmup-mode request --save-output
 ```
 
 For Hunyuan3D, treat `Hunyuan3DShapeDenoisingStage` as the primary latency
@@ -275,7 +275,7 @@ drive DiT optimization decisions.
 
 ```bash
 sglang generate --model-path <MODEL> \
-  --enable-torch-compile --warmup \
+  --enable-torch-compile --warmup-mode request \
   --dit-layerwise-offload --dit-offload-prefetch-size 0.1 \
   --text-encoder-cpu-offload true --vae-cpu-offload true \
   --prompt "..." --save-output
@@ -287,7 +287,7 @@ sglang generate --model-path <MODEL> \
 SGLANG_CACHE_DIT_ENABLED=true sglang generate --model-path <MODEL> \
   --attention-backend sage_attn \
   --dit-layerwise-offload false \
-  --enable-torch-compile --warmup \
+  --enable-torch-compile --warmup-mode request \
   --prompt "..." --save-output
 ```
 
@@ -314,22 +314,22 @@ Use these as first commands to benchmark, not as universal winners.
 | Model family | First performance shape | Starting flags | Notes |
 |---|---|---|---|
 | MiniMax-H3 | 1344x768 resolved canvas, 5 seconds / 124 frames at 24 fps, 50 joint video/audio steps | H200: `--num-gpus 4 --ulysses-degree 4 --performance-mode speed --enable-torch-compile false`; H100: TP2 + Ulysses2 | Root ID plus `--model-variant fl2va` for T2VA/FL2VA or `ref2va` for Ref2VA. Ulysses only; no Ring/CFG/SageAttention. Preserve tiled video-VAE decode. Profile joint denoise, video VAE, audio VAE/vocoder, encoder, and collectives separately. |
-| FLUX.1 / FLUX.2 image | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup --dit-layerwise-offload false` | `black-forest-labs/FLUX.*` repos are gated; for FP8/NVFP4 use validated `--transformer-path` or `--transformer-weights-path` flows from the quant skill. |
-| FLUX.2 Klein / Klein Base | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup --dit-layerwise-offload false` | Current registry has `black-forest-labs/FLUX.2-klein-4B`, `FLUX.2-klein-9B`, and base variants. Klein is step-distilled; Klein Base is not. |
-| Qwen-Image / Qwen-Image-Edit | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup`; optionally native `SGLANG_CACHE_DIT_ENABLED=true` | Cache-DiT is lossy. For edit tasks, keep reference image, seed, and output size fixed. |
-| Z-Image / Z-Image-Turbo | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup` | Keep base Z-Image separate from Turbo: base uses 50-step CFG defaults, Turbo uses 9-step zero-CFG defaults. Mainline has bf16-native Triton RMSNorm scale and tanh-residual fusions. |
+| FLUX.1 / FLUX.2 image | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup-mode request --dit-layerwise-offload false` | `black-forest-labs/FLUX.*` repos are gated; for FP8/NVFP4 use validated `--transformer-path` or `--transformer-weights-path` flows from the quant skill. |
+| FLUX.2 Klein / Klein Base | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup-mode request --dit-layerwise-offload false` | Current registry has `black-forest-labs/FLUX.2-klein-4B`, `FLUX.2-klein-9B`, and base variants. Klein is step-distilled; Klein Base is not. |
+| Qwen-Image / Qwen-Image-Edit | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup-mode request`; optionally native `SGLANG_CACHE_DIT_ENABLED=true` | Cache-DiT is lossy. For edit tasks, keep reference image, seed, and output size fixed. |
+| Z-Image / Z-Image-Turbo | 1024x1024, runtime-default steps/guidance, 1 GPU | `--enable-torch-compile --warmup-mode request` | Keep base Z-Image separate from Turbo: base uses 50-step CFG defaults, Turbo uses 9-step zero-CFG defaults. Mainline has bf16-native Triton RMSNorm scale and tanh-residual fusions. |
 | Wan2.2 A14B T2V/I2V | 1280x720, 81 frames | Nightly: `--num-gpus 4 --enable-cfg-parallel --ulysses-degree 2 --text-encoder-cpu-offload --pin-cpu-memory` | For lowest latency, also benchmark pure Ulysses on the same GPUs. |
-| Wan2.2 TI2V 5B | 1280x720, 81 frames, 1 GPU | `--enable-torch-compile --warmup` | Keep the input image and motion prompt fixed when comparing sparse attention or Cache-DiT. |
-| Wan2.1 / FastWan / TurboWan variants | 480p or 720p video, family defaults | `--enable-torch-compile --warmup`; add `--ulysses-degree` / CFG parallel only after measuring | Current registry includes Wan2.1, FastWan2.1, FastWan2.2 TI2V, TurboWan2.1, TurboWan2.2 I2V, and Wan2.1-Fun InP. Use the compatibility matrix and benchmark presets before choosing topology. |
-| Cosmos3 Nano / Super | T2I: 1024x1024 with `--num-frames 1`; T2V/I2V: 480p/720p video | `SGLANG_DISABLE_COSMOS3_GUARDRAILS=1` for benchmark isolation; `--enable-torch-compile --warmup` | One checkpoint serves T2I/T2V/I2V. Mode is request-driven: `num_frames == 1` means T2I, `--image-path` means I2V. |
-| Ideogram 4 FP8/NVFP4 | 1024x1024, native preset defaults | `--enable-torch-compile --warmup` | Do not set `--num-inference-steps` or `--guidance-scale` directly unless you also update the Ideogram preset; sampling params derive them from `preset`. |
-| ERNIE-Image / GLM-Image / SANA / SD3 | 1024-class image, family defaults | `--enable-torch-compile --warmup`; disable offload only after checking VRAM | Treat these as current native image families. Start with benchmark/profile presets for ERNIE, GLM, and SANA; use registry/config defaults for SD3 unless you add a new preset. |
-| LTX-2 / LTX-2.3 | 768x512 or HQ 1920x1088, 121 frames | `--pipeline-class-name LTX2TwoStagePipeline --enable-torch-compile --warmup`; HQ uses `LTX2TwoStageHQPipeline` | Use benchmark/profile presets for nightly alignment, one-stage, high-resolution stress, and HQ. Device mode choices are `original` and `resident`; `resident` is fastest but uses more VRAM. `snapshot` is a deprecated alias for `original`, so do not use it in new commands. |
-| HunyuanVideo | 848x480 or 720p class video | `--text-encoder-cpu-offload --pin-cpu-memory --enable-torch-compile --warmup` | Check VAE decode separately. GroupNorm+SiLU is default-eligible in mainline when wrapper guards pass; use `bench_group_norm_silu.py` when VAE residual blocks are hot. |
-| JoyAI-Image-Edit | 1024-class TI2I, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Newly supported image-edit path. Keep the input image, prompt, seed, and output size fixed; 2-GPU CFG parallel is the validated H100 starting point. |
-| FireRed-Image-Edit 1.0 / 1.1 | 1024x1024 image edit, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Uses the native `QwenImageEditPlusPipeline` path. 2-GPU CFG parallel is the validated H100 starting point; benchmark 1.0 and 1.1 separately because checkpoint differences can change denoise latency. |
-| Hunyuan3D-2 shape | Shape generation, 50 steps, guidance 5.0 | `--backend=sglang --enable-torch-compile --warmup --dit-layerwise-offload false --dit-cpu-offload false` | Focus on `Hunyuan3DShapeDenoisingStage`; keep mesh export/paint timings separate from denoise. |
-| MOVA / Helios / LingBot World | Use the benchmark/profile presets or server test cases first | `--enable-torch-compile --warmup`; pin offload and topology flags explicitly | These video/realtime families have model-specific stages and condition handling. Keep prompt/image/action inputs fixed and prefer perf dumps over wall time alone. |
+| Wan2.2 TI2V 5B | 1280x720, 81 frames, 1 GPU | `--enable-torch-compile --warmup-mode request` | Keep the input image and motion prompt fixed when comparing sparse attention or Cache-DiT. |
+| Wan2.1 / FastWan / TurboWan variants | 480p or 720p video, family defaults | `--enable-torch-compile --warmup-mode request`; add `--ulysses-degree` / CFG parallel only after measuring | Current registry includes Wan2.1, FastWan2.1, FastWan2.2 TI2V, TurboWan2.1, TurboWan2.2 I2V, and Wan2.1-Fun InP. Use the compatibility matrix and benchmark presets before choosing topology. |
+| Cosmos3 Nano / Super | T2I: 1024x1024 with `--num-frames 1`; T2V/I2V: 480p/720p video | `SGLANG_DISABLE_COSMOS3_GUARDRAILS=1` for benchmark isolation; `--enable-torch-compile --warmup-mode request` | One checkpoint serves T2I/T2V/I2V. Mode is request-driven: `num_frames == 1` means T2I, `--image-path` means I2V. |
+| Ideogram 4 FP8/NVFP4 | 1024x1024, native preset defaults | `--enable-torch-compile --warmup-mode request` | Do not set `--num-inference-steps` or `--guidance-scale` directly unless you also update the Ideogram preset; sampling params derive them from `preset`. |
+| ERNIE-Image / GLM-Image / SANA / SD3 | 1024-class image, family defaults | `--enable-torch-compile --warmup-mode request`; disable offload only after checking VRAM | Treat these as current native image families. Start with benchmark/profile presets for ERNIE, GLM, and SANA; use registry/config defaults for SD3 unless you add a new preset. |
+| LTX-2 / LTX-2.3 | 768x512 or HQ 1920x1088, 121 frames | `--pipeline-class-name LTX2TwoStagePipeline --enable-torch-compile --warmup-mode request`; HQ uses `LTX2TwoStageHQPipeline` | Use benchmark/profile presets for nightly alignment, one-stage, high-resolution stress, and HQ. Device mode choices are `original` and `resident`; `resident` is fastest but uses more VRAM. `snapshot` is a deprecated alias for `original`, so do not use it in new commands. |
+| HunyuanVideo | 848x480 or 720p class video | `--text-encoder-cpu-offload --pin-cpu-memory --enable-torch-compile --warmup-mode request` | Check VAE decode separately. GroupNorm+SiLU is default-eligible in mainline when wrapper guards pass; use `bench_group_norm_silu.py` when VAE residual blocks are hot. |
+| JoyAI-Image-Edit | 1024-class TI2I, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup-mode request --dit-layerwise-offload false --dit-cpu-offload false` | Newly supported image-edit path. Keep the input image, prompt, seed, and output size fixed; 2-GPU CFG parallel is the validated H100 starting point. |
+| FireRed-Image-Edit 1.0 / 1.1 | 1024x1024 image edit, 40 steps, guidance 4.0 | `--backend=sglang --num-gpus 2 --enable-cfg-parallel --ulysses-degree 1 --enable-torch-compile --warmup-mode request --dit-layerwise-offload false --dit-cpu-offload false` | Uses the native `QwenImageEditPlusPipeline` path. 2-GPU CFG parallel is the validated H100 starting point; benchmark 1.0 and 1.1 separately because checkpoint differences can change denoise latency. |
+| Hunyuan3D-2 shape | Shape generation, 50 steps, guidance 5.0 | `--backend=sglang --enable-torch-compile --warmup-mode request --dit-layerwise-offload false --dit-cpu-offload false` | Focus on `Hunyuan3DShapeDenoisingStage`; keep mesh export/paint timings separate from denoise. |
+| MOVA / Helios / LingBot World | Use the benchmark/profile presets or server test cases first | `--enable-torch-compile --warmup-mode request`; pin offload and topology flags explicitly | These video/realtime families have model-specific stages and condition handling. Keep prompt/image/action inputs fixed and prefer perf dumps over wall time alone. |
 
 ## Historical PR Watchlist
 
@@ -343,7 +343,7 @@ about whether the work has merged:
 
 ## Tips
 
-- **Benchmarking**: always use `--warmup` and look for the line ending with `(with warmup excluded)` for accurate timing.
+- **Benchmarking**: always use `--warmup-mode request` and look for the line ending with `(with warmup excluded)` for accurate timing.
 - **Perf dump**: use `--perf-dump-path result.json` to save structured metrics, then compare with `python python/sglang/multimodal_gen/benchmarks/compare_perf.py baseline.json result.json`.
 - **Offload tuning**: after the first request, the runtime logs peak GPU memory and which components could stay resident. Use this to decide which `--*-cpu-offload` flags to disable.
 - **Backend selection**: `--backend sglang` (default, auto-detected) enables native optimizations (fused kernels, SP, native Cache-DiT env knobs, etc.). `--backend diffusers` falls back to Diffusers pipelines and is the path that accepts `--cache-dit-config` plus diffusers attention backend names.

python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py

@@ -472,7 +472,7 @@ class DiffGenerator:
     def _log_summary(self, results: list[GenerationResult]) -> None:
         if not results:
             return
-        if self.server_args.warmup:
+        if self.server_args.warmup_mode != "off":
             total_duration_ms = results[0].metrics.get("total_duration_ms", 0)
             logger.info(
                 f"Warmed-up request processed in {GREEN}%.2f{RESET} seconds (with warmup excluded)",

python/sglang/multimodal_gen/runtime/entrypoints/http_server.py

@@ -118,7 +118,7 @@ async def lifespan(app: FastAPI):
     # 2. Start the ZMQ Broker in the background to handle offline requests
     broker_task = asyncio.create_task(run_zeromq_broker(server_args))
     warmup_task = None
-    if server_args.server_warmup:
+    if server_args.warmup_mode == "server":
         warmup_task = asyncio.create_task(
             _run_server_warmup_after_http_ready(server_args, warmup_done)
         )

python/sglang/multimodal_gen/runtime/launch_server.py

@@ -411,8 +411,7 @@ def launch_pool_disagg_server(
                 "pool_work_endpoint": work_eps[inst_idx],
                 "pool_result_endpoint": result_ep,
                 "num_gpus": num_role_gpus,
-                "warmup": role_type == RoleType.ENCODER,
-                "server_warmup": False,
+                "warmup_mode": "request" if role_type == RoleType.ENCODER else "off",
                 "scheduler_port": find_port(port_cursor),
                 "master_port": find_port(port_cursor + 100),
                 # Per-role parallelism (None = auto-derive from num_gpus)
@@ -692,8 +691,7 @@ def launch_disagg_role(server_args: ServerArgs):
         "disagg_mode": True,
         "pool_work_endpoint": work_endpoint,
         "pool_result_endpoint": result_endpoint,
-        "warmup": role_type == RoleType.ENCODER,
-        "server_warmup": False,
+        "warmup_mode": "request" if role_type == RoleType.ENCODER else "off",
         "scheduler_port": internal_scheduler_port,
         # Per-role parallelism (None = auto-derive from num_gpus)
         "tp_size": role_par["tp_size"],

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

@@ -358,7 +358,7 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
         if (
             not args.enable_torch_compile
             or not args.offload_during_compile
-            or not args.warmup
+            or args.warmup_mode == "off"
             or not self._owns_compile_warmup_lifecycle()
             or args.use_fsdp_inference
             or self._cache_dit_requested()

python/sglang/multimodal_gen/runtime/server_args/disagg.py

@@ -234,9 +234,3 @@ class DisaggServerArgsMixin:
             default=cls.decoder_sp,
             help="Sequence parallelism for decoder role.",
         )
-        parser.add_argument(
-            "--decoder-tp",
-            type=int,
-            default=cls.decoder_tp,
-            help="Deprecated alias for --decoder-sp.",
-        )

python/sglang/multimodal_gen/runtime/server_args/server_args.py

@@ -315,21 +315,15 @@ class ServerArgs(DisaggServerArgsMixin):
     # NVTX profiling
     enable_layerwise_nvtx_marker: bool = False
 
-    # warmup
-    # `warmup_mode` is the canonical knob: one of WARMUP_MODES
+    # Warmup is controlled by the canonical `warmup_mode` knob: one of WARMUP_MODES.
     #   - "off":     no warmup.
     #   - "server":  server-based warmup — a synthetic request right after the
     #                server is ready, before real traffic
     #   - "request": request-based warmup — warm on the first real request(s).
-    #                This is a BENCHMARK aid
-    # existing consumers keep working) and as deprecated CLI aliases. None means
-    # "derive the mode from the legacy booleans"; _adjust_warmup resolves it.
+    #                This is a BENCHMARK aid.
+    # None is resolved by _adjust_warmup from the selected runtime features.
     warmup_mode: str | None = None
 
-    # deprecated: warmup and server_warmup
-    warmup: bool = False
-    server_warmup: bool = False
-
     warmup_resolutions: list[str] = None
     warmup_steps: int = 1
 
@@ -411,7 +405,6 @@ class ServerArgs(DisaggServerArgsMixin):
     denoiser_ulysses: int | None = None
     denoiser_ring: int | None = None
     decoder_sp: int | None = None
-    decoder_tp: int | None = None
     pool_work_endpoint: str | None = None
     pool_result_endpoint: str | None = None
     pool_control_endpoint: str | None = None
@@ -457,7 +450,6 @@ class ServerArgs(DisaggServerArgsMixin):
         """set defaults and normalize values."""
         auto_tuner = ServerArgsAutoTuner(self)
         auto_tuner.adjust_based_on_performance_mode()
-        self._adjust_disagg_parallelism_aliases()
         if auto_tuner.could_override_server_args():
             self._adjust_offload()
             auto_tuner.maybe_adjust_auto_default_layerwise_offload()
@@ -480,21 +472,6 @@ class ServerArgs(DisaggServerArgsMixin):
         auto_tuner.finalize_auto_flags()
         self.adjust_pipeline_config()
 
-    def _adjust_disagg_parallelism_aliases(self):
-        if self.decoder_tp is None:
-            return
-        if self.decoder_sp is not None and self.decoder_sp != self.decoder_tp:
-            raise ValueError(
-                "decoder_tp is deprecated in favor of decoder_sp; "
-                "please set only one of them or keep the same value."
-            )
-        if self.decoder_sp is None:
-            logger.warning(
-                "decoder_tp is deprecated and is treated as decoder_sp for "
-                "decoder/VAE parallel decode. Please use decoder_sp instead."
-            )
-            self.decoder_sp = self.decoder_tp
-
     def _validate_parameters(self):
         """check consistency and raise errors for invalid configs"""
         self._validate_pipeline()
@@ -889,64 +866,37 @@ class ServerArgs(DisaggServerArgsMixin):
         return None, None
 
     def _adjust_warmup(self):
-        #   --warmup-mode > --warmup/--server-warmup
-        mode_explicit = self.is_arg_explicitly_set("warmup_mode")
-        legacy_explicit = self.is_arg_explicitly_set(
-            "warmup"
-        ) or self.is_arg_explicitly_set("server_warmup")
-        if self.warmup_mode is not None:
-            if self.warmup_mode not in WARMUP_MODES:
-                raise ValueError(
-                    f"Invalid --warmup-mode {self.warmup_mode!r}; "
-                    f"expected one of {WARMUP_MODES}."
-                )
-            if mode_explicit and legacy_explicit:
-                logger.warning(
-                    "Both --warmup-mode and the deprecated --warmup/--server-warmup "
-                    "were set; --warmup-mode=%s takes precedence.",
-                    self.warmup_mode,
-                )
-            if mode_explicit or not legacy_explicit:
-                self.warmup = self.warmup_mode != "off"
-                self.server_warmup = self.warmup_mode == "server"
-            elif self.warmup:
-                self.server_warmup = self.server_warmup or self.warmup_mode == "server"
+        if self.warmup_mode is not None and self.warmup_mode not in WARMUP_MODES:
+            raise ValueError(
+                f"Invalid --warmup-mode {self.warmup_mode!r}; "
+                f"expected one of {WARMUP_MODES}."
+            )
 
-        # Explicit resolutions imply warmup is on (request-based).
-        if self.warmup_resolutions is not None:
-            self.warmup = True
-
-        if (
-            self.enable_torch_compile
-            and self.warmup_mode is None
-            and not mode_explicit
-            and not legacy_explicit
-        ):
-            self.warmup = True
-            self.server_warmup = True
+        if self.enable_torch_compile and self.warmup_mode is None:
+            self.warmup_mode = "server"
             logger.info(
                 "Automatically enabled server warmup for torch.compile so first "
                 "real requests do not pay compile latency. Set --warmup-mode off "
                 "to disable this behavior."
             )
 
+        # Explicit resolutions need a request path unless an existing server
+        # default already supplies the synthetic startup request.
+        if self.warmup_resolutions is not None and self.warmup_mode in (None, "off"):
+            self.warmup_mode = "request"
+
         # BCG captures every graph during a synthetic warmup forward at startup
-        # so that serving never records a fresh graph. That requires
-        # server-based warmup (a real warmup request issued at startup), not
-        # request-based warmup which runs no forward until the first request.
+        # so serving never records a fresh graph.
         if self.enable_breakable_cuda_graph and self.disagg_role == RoleType.MONOLITHIC:
-            self.warmup = True
-            self.server_warmup = True
+            self.warmup_mode = "server"
 
-        if self.disagg_role != RoleType.MONOLITHIC:
-            self.server_warmup = False
+        # Disaggregated roles do not host the HTTP startup request. Preserve
+        # warmup intent, but schedule it on the first request instead.
+        if self.disagg_role != RoleType.MONOLITHIC and self.warmup_mode == "server":
+            self.warmup_mode = "request"
 
-        if not self.warmup:
-            self.server_warmup = False
-
-        self.warmup_mode = (
-            "off" if not self.warmup else "server" if self.server_warmup else "request"
-        )
+        if self.warmup_mode is None:
+            self.warmup_mode = "off"
 
     @staticmethod
     def _require_port(port: int, name: str) -> None:
@@ -1622,28 +1572,16 @@ class ServerArgs(DisaggServerArgsMixin):
             choices=list(WARMUP_MODES),
             default=ServerArgs.warmup_mode,
             help=(
-                "Warmup mode (canonical knob). One of: "
-                "`off` (no warmup); `request` (request-based: warm on real "
-                "incoming requests); `server` (server-based: a synthetic warmup "
-                "request right after the server is ready, before traffic). "
-                "Takes precedence over the deprecated --warmup/--server-warmup. "
-                "`sglang serve` defaults to `server`; other entrypoints default "
+                "Warmup mode. One of: `off` (no warmup); `request` "
+                "(request-based: warm on real incoming requests); `server` "
+                "(server-based: a synthetic warmup request right after the server "
+                "is ready, before traffic). `sglang serve` defaults to `server`; "
+                "other entrypoints default "
                 "to request-based when warmup is enabled. When enabled, look for "
                 "the line ending with `(with warmup excluded)` for actual "
                 "processing time."
             ),
         )
-        parser.add_argument(
-            "--warmup",
-            action=StoreBoolean,
-            default=ServerArgs.warmup,
-            help=(
-                "[DEPRECATED: use --warmup-mode] Perform warmup before normal "
-                "traffic. Maps to --warmup-mode request (or server, combined "
-                "with --server-warmup). Recommended when benchmarking for fair "
-                "comparison and best performance."
-            ),
-        )
         parser.add_argument(
             "--warmup-resolutions",
             type=str,
@@ -1657,16 +1595,6 @@ class ServerArgs(DisaggServerArgsMixin):
             default=ServerArgs.warmup_steps,
             help="The number of warmup steps to perform for each resolution.",
         )
-        parser.add_argument(
-            "--server-warmup",
-            action=StoreBoolean,
-            default=ServerArgs.server_warmup,
-            help=(
-                "[DEPRECATED: use --warmup-mode server] Send a synthetic warmup "
-                "request after the server is ready (server-based warmup)."
-            ),
-        )
-
         # layerwise offload
         parser.add_argument(
             "--dit-cpu-offload",
@@ -2238,6 +2166,7 @@ class ServerArgs(DisaggServerArgsMixin):
     @classmethod
     def from_dict(cls, kwargs: dict[str, Any]) -> "ServerArgs":
         """Create a ServerArgs object from a dictionary."""
+        cls._reject_retired_args(kwargs)
         attrs = [attr.name for attr in dataclasses.fields(cls) if attr.init]
         server_args_kwargs: dict[str, Any] = {}
         explicit_arg_names = kwargs.get("_explicit_arg_names")
@@ -2264,6 +2193,20 @@ class ServerArgs(DisaggServerArgsMixin):
 
         return cls(**server_args_kwargs)
 
+    @staticmethod
+    def _reject_retired_args(kwargs: dict[str, Any]) -> None:
+        retired_args = {
+            "decoder_tp": "decoder_sp for decoder/VAE parallel decode",
+            "warmup": "warmup_mode=request or warmup_mode=off",
+            "server_warmup": "warmup_mode=server or warmup_mode=off",
+        }
+        removed = [name for name in retired_args if name in kwargs]
+        if removed:
+            replacements = "; ".join(
+                f"{name} -> {retired_args[name]}" for name in removed
+            )
+            raise ValueError(f"Removed server argument(s): {replacements}")
+
     @staticmethod
     def load_config_file(config_file: str) -> dict[str, Any]:
         """Load a config file."""
@@ -2285,6 +2228,7 @@ class ServerArgs(DisaggServerArgsMixin):
 
     @classmethod
     def from_kwargs(cls, **kwargs: Any) -> "ServerArgs":
+        cls._reject_retired_args(kwargs)
         explicit_arg_names = set(kwargs)
 
         # Convert backend string to enum if necessary

python/sglang/multimodal_gen/runtime/server_warmup.py

@@ -69,7 +69,7 @@ def should_return_warmup_result(req_or_group: Any) -> bool:
 
 
 def should_run_server_warmup(server_args: ServerArgs) -> bool:
-    return server_args.warmup and server_args.server_warmup
+    return server_args.warmup_mode == "server"
 
 
 def is_realtime_serving(server_args: ServerArgs) -> bool:
@@ -95,7 +95,7 @@ def should_run_synthetic_server_warmup(server_args: ServerArgs) -> bool:
 
 def should_run_explicit_client_warmup(server_args: ServerArgs) -> bool:
     return (
-        server_args.warmup
+        server_args.warmup_mode != "off"
         and server_args.warmup_resolutions is not None
         and supports_synthetic_warmup(server_args)
     )
@@ -298,10 +298,9 @@ class SchedulerWarmupMixin:
     ) -> list[tuple[bytes, Any]]:
         if (
             self.req_based_warmup_scheduled
-            or not self.server_args.warmup
+            or self.server_args.warmup_mode != "request"
             or not recv_reqs
             or self.server_args.warmup_resolutions is not None
-            or self.server_args.server_warmup
         ):
             return recv_reqs
 

python/sglang/multimodal_gen/test/scripts/gen_perf_baselines.py

@@ -68,8 +68,8 @@ def _build_server_extra_args(case: DiffusionTestCase) -> str:
     if server_args.lora_path:
         a += f" --lora-path {server_args.lora_path}"
 
-    # default warmup
-    a += " --warmup"
+    # request-based warmup keeps the first measured generation out of the baseline
+    a += " --warmup-mode request"
 
     for extra_arg in server_args.extras:
         a += f" {extra_arg}"

python/sglang/multimodal_gen/test/server/gpu_cases.py

@@ -457,7 +457,7 @@ ONE_GPU_CASES: list[DiffusionTestCase] = [
             modality="video",
             num_gpus=1,
             extras=[
-                "--pipeline-class-name LingBotWorldCausalDMDPipeline --warmup false"
+                "--pipeline-class-name LingBotWorldCausalDMDPipeline --warmup-mode off"
             ],
             text_encoder_cpu_offload=True,
         ),

python/sglang/multimodal_gen/test/single_test_file/test_pi05_e2e.py

@@ -77,7 +77,7 @@ def pi05_generator():
     kwargs = {
         "model_path": _MODEL_PATH,
         "num_gpus": num_gpus,
-        "warmup": False,
+        "warmup_mode": "off",
         "trust_remote_code": False,
     }
     if num_gpus > 1:

python/sglang/multimodal_gen/test/unit/test_cfg_parallel_warmup.py

@@ -64,12 +64,11 @@ def _make_bare_scheduler(enable_cfg_parallel: bool) -> Scheduler:
     scheduler = object.__new__(Scheduler)
 
     server_args = MagicMock()
-    server_args.warmup = True
+    server_args.warmup_mode = "request"
     server_args.warmup_steps = 1
     server_args.warmup_resolutions = ["512x512"]
     server_args.enable_cfg_parallel = enable_cfg_parallel
     server_args.enable_torch_compile = False
-    server_args.server_warmup = False
     server_args.is_arg_explicitly_set.return_value = False
 
     task_type = MagicMock()
@@ -235,7 +234,7 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
     def test_req_based_warmup_remains_explicit_legacy_entry(self):
         scheduler = _make_bare_scheduler(enable_cfg_parallel=False)
         scheduler.server_args.warmup_resolutions = None
-        scheduler.server_args.server_warmup = False
+        scheduler.server_args.warmup_mode = "request"
 
         req = _make_generation_req()
         recv_reqs = [(b"0", req)]
@@ -253,7 +252,7 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
     def test_req_based_warmup_skips_default_server_warmup_path(self):
         scheduler = _make_bare_scheduler(enable_cfg_parallel=False)
         scheduler.server_args.warmup_resolutions = None
-        scheduler.server_args.server_warmup = True
+        scheduler.server_args.warmup_mode = "server"
 
         recv_reqs = [(b"0", _make_generation_req())]
         processed = scheduler.process_received_reqs_with_req_based_warmup(recv_reqs)
@@ -265,7 +264,7 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
     def test_diff_generator_runs_explicit_warmup_through_scheduler_client(self):
         generator = object.__new__(DiffGenerator)
         server_args = MagicMock()
-        server_args.warmup = True
+        server_args.warmup_mode = "request"
         server_args.warmup_resolutions = ["832x480"]
         server_args.warmup_steps = 1
         server_args.enable_cfg_parallel = False
@@ -693,8 +692,7 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
 
     def test_action_pipeline_disables_synthetic_warmup(self):
         server_args = MagicMock()
-        server_args.warmup = True
-        server_args.server_warmup = True
+        server_args.warmup_mode = "server"
         server_args.warmup_resolutions = ["512x512"]
         server_args.pipeline_config.task_type = ModelTaskType.VLA_ACTION
 
@@ -704,8 +702,7 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
 
     def test_mesh_pipeline_builds_image_conditioned_warmup(self):
         server_args = MagicMock()
-        server_args.warmup = True
-        server_args.server_warmup = True
+        server_args.warmup_mode = "server"
         server_args.warmup_steps = 1
         server_args.warmup_resolutions = None
         server_args.enable_cfg_parallel = False

python/sglang/multimodal_gen/test/unit/test_server_args.py

@@ -421,12 +421,10 @@ class TestServerArgsPathExpansion(unittest.TestCase):
             execute_serve_cmd(args, unknown_args)
 
         server_args = dispatch_launch.call_args.args[0]
-        self.assertTrue(server_args.warmup)
-        self.assertTrue(server_args.server_warmup)
-        self.assertFalse(server_args.is_arg_explicitly_set("warmup"))
-        self.assertFalse(server_args.is_arg_explicitly_set("server_warmup"))
+        self.assertEqual(server_args.warmup_mode, "server")
+        self.assertFalse(server_args.is_arg_explicitly_set("warmup_mode"))
 
-    def test_serve_cli_preserves_explicit_warmup_false(self):
+    def test_serve_cli_preserves_explicit_warmup_mode_off(self):
         from sglang.multimodal_gen.runtime.entrypoints.cli.serve import (
             add_multimodal_gen_serve_args,
             execute_serve_cmd,
@@ -437,8 +435,8 @@ class TestServerArgsPathExpansion(unittest.TestCase):
         argv = [
             "--model-path",
             "/fake",
-            "--warmup",
-            "false",
+            "--warmup-mode",
+            "off",
         ]
 
         with (
@@ -454,18 +452,17 @@ class TestServerArgsPathExpansion(unittest.TestCase):
             execute_serve_cmd(args, unknown_args)
 
         server_args = dispatch_launch.call_args.args[0]
-        self.assertFalse(server_args.warmup)
-        self.assertFalse(server_args.server_warmup)
-        self.assertTrue(server_args.is_arg_explicitly_set("warmup"))
+        self.assertEqual(server_args.warmup_mode, "off")
+        self.assertTrue(server_args.is_arg_explicitly_set("warmup_mode"))
 
-    def test_serve_cli_preserves_config_warmup_false(self):
+    def test_serve_cli_preserves_config_warmup_mode_off(self):
         from sglang.multimodal_gen.runtime.entrypoints.cli.serve import (
             add_multimodal_gen_serve_args,
             execute_serve_cmd,
         )
 
         with tempfile.NamedTemporaryFile("w", suffix=".json") as config_file:
-            json.dump({"model_path": "/fake", "warmup": False}, config_file)
+            json.dump({"model_path": "/fake", "warmup_mode": "off"}, config_file)
             config_file.flush()
 
             parser = FlexibleArgumentParser()
@@ -490,9 +487,18 @@ class TestServerArgsPathExpansion(unittest.TestCase):
                 execute_serve_cmd(args, unknown_args)
 
         server_args = dispatch_launch.call_args.args[0]
-        self.assertFalse(server_args.warmup)
-        self.assertFalse(server_args.server_warmup)
-        self.assertTrue(server_args.is_arg_explicitly_set("warmup"))
+        self.assertEqual(server_args.warmup_mode, "off")
+        self.assertTrue(server_args.is_arg_explicitly_set("warmup_mode"))
+
+    def test_retired_warmup_config_is_rejected(self):
+        with self.assertRaisesRegex(ValueError, "warmup.*warmup_mode"):
+            _from_dict_without_model_resolution(
+                {"model_path": "/fake", "warmup": False}
+            )
+
+    def test_retired_warmup_kwargs_are_rejected(self):
+        with self.assertRaisesRegex(ValueError, "warmup.*warmup_mode"):
+            ServerArgs.from_kwargs(model_path="/fake", warmup=False)
 
     def test_disagg_role_disables_server_warmup(self):
         with patch.object(
@@ -501,156 +507,82 @@ class TestServerArgsPathExpansion(unittest.TestCase):
             server_args = ServerArgs.from_dict(
                 {
                     "model_path": "/fake",
-                    "warmup": True,
-                    "server_warmup": True,
+                    "warmup_mode": "server",
                     "disagg_role": "server",
                 }
             )
 
-        self.assertTrue(server_args.warmup)
-        self.assertFalse(server_args.server_warmup)
+        self.assertEqual(server_args.warmup_mode, "request")
 
 
 class TestWarmupModeNormalization(unittest.TestCase):
-    """`_adjust_warmup` resolves the canonical warmup_mode and its derived booleans."""
+    """`_adjust_warmup` resolves the canonical warmup mode."""
 
     def _resolve(
         self,
         *,
         warmup_mode=None,
-        warmup=False,
-        server_warmup=False,
         warmup_resolutions=None,
         enable_torch_compile=False,
+        enable_breakable_cuda_graph=False,
         disagg_role=None,
-        explicit=(),
     ):
         from sglang.multimodal_gen.runtime.disaggregation.roles import RoleType
 
         sa = ServerArgs.__new__(ServerArgs)
         sa.warmup_mode = warmup_mode
-        sa.warmup = warmup
-        sa.server_warmup = server_warmup
         sa.warmup_resolutions = warmup_resolutions
         sa.enable_torch_compile = enable_torch_compile
+        sa.enable_breakable_cuda_graph = enable_breakable_cuda_graph
         sa.disagg_role = RoleType.MONOLITHIC if disagg_role is None else disagg_role
-        sa._explicit_arg_names = set(explicit)
         sa._adjust_warmup()
         return sa
 
     def test_explicit_mode_off_disables_all(self):
-        sa = self._resolve(warmup_mode="off", explicit=("warmup_mode",))
+        sa = self._resolve(warmup_mode="off")
         self.assertEqual(sa.warmup_mode, "off")
-        self.assertFalse(sa.warmup)
-        self.assertFalse(sa.server_warmup)
 
     def test_explicit_mode_request(self):
-        sa = self._resolve(warmup_mode="request", explicit=("warmup_mode",))
+        sa = self._resolve(warmup_mode="request")
         self.assertEqual(sa.warmup_mode, "request")
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
 
     def test_explicit_mode_server(self):
-        sa = self._resolve(warmup_mode="server", explicit=("warmup_mode",))
-        self.assertEqual(sa.warmup_mode, "server")
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
-
-    def test_explicit_mode_overrides_explicit_legacy(self):
-        sa = self._resolve(
-            warmup_mode="request",
-            warmup=True,
-            server_warmup=True,
-            explicit=("warmup_mode", "warmup", "server_warmup"),
-        )
-        self.assertEqual(sa.warmup_mode, "request")
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
-
-    def test_explicit_legacy_false_beats_defaulted_mode(self):
-        # serve defaults warmup_mode="server" (not explicit); `--warmup false` wins.
-        sa = self._resolve(
-            warmup_mode="server",
-            warmup=False,
-            server_warmup=False,
-            explicit=("warmup",),
-        )
-        self.assertEqual(sa.warmup_mode, "off")
-        self.assertFalse(sa.warmup)
-        self.assertFalse(sa.server_warmup)
-
-    def test_defaulted_mode_applies_without_legacy_flags(self):
-        # bare `sglang serve`: warmup_mode="server" defaulted, no legacy override.
         sa = self._resolve(warmup_mode="server")
         self.assertEqual(sa.warmup_mode, "server")
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
 
-    def test_legacy_only_maps_to_request(self):
-        sa = self._resolve(warmup_mode=None, warmup=True, explicit=("warmup",))
-        self.assertEqual(sa.warmup_mode, "request")
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
+    def test_defaulted_mode_applies_without_legacy_flags(self):
+        # Bare `sglang serve` defaults to server-based warmup.
+        sa = self._resolve(warmup_mode="server")
+        self.assertEqual(sa.warmup_mode, "server")
 
     def test_resolutions_force_warmup_on(self):
         sa = self._resolve(
             warmup_mode="off",
             warmup_resolutions=["512x512"],
-            explicit=("warmup_mode",),
         )
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
         self.assertEqual(sa.warmup_mode, "request")
 
     def test_torch_compile_defaults_to_server_warmup(self):
         sa = self._resolve(enable_torch_compile=True)
 
         self.assertEqual(sa.warmup_mode, "server")
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
-
-    def test_legacy_warmup_on_uses_defaulted_server_mode(self):
-        # `serve --warmup` (legacy ON, mode defaulted to "server" but not
-        # explicit) must resolve to server-based warmup, not silently downgrade
-        # to request mode.
-        sa = self._resolve(warmup_mode="server", warmup=True, explicit=("warmup",))
-
-        self.assertEqual(sa.warmup_mode, "server")
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
 
     def test_torch_compile_respects_explicit_warmup_off(self):
         sa = self._resolve(
             warmup_mode="off",
             enable_torch_compile=True,
-            explicit=("warmup_mode",),
         )
         self.assertEqual(sa.warmup_mode, "off")
-        self.assertFalse(sa.warmup)
-        self.assertFalse(sa.server_warmup)
 
     def test_torch_compile_uses_server_warmup_for_explicit_resolutions(self):
         sa = self._resolve(
             warmup_resolutions=["1024x1024"],
             enable_torch_compile=True,
-            explicit=("warmup_resolutions",),
         )
         self.assertEqual(sa.warmup_mode, "server")
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
 
-    def test_legacy_warmup_with_resolutions_runs_server_warmup(self):
-        # Dead-zone regression: `serve --warmup --warmup-resolutions X` must run
-        # server-based (synthetic) warmup, not end up with no warmup at all
-        # (request-based warmup bails out when warmup_resolutions is set).
-        sa = self._resolve(
-            warmup_mode="server",
-            warmup=True,
-            warmup_resolutions=["1024x1024"],
-            explicit=("warmup",),
-        )
-        self.assertTrue(sa.warmup)
-        self.assertTrue(sa.server_warmup)
+    def test_breakable_cuda_graph_forces_server_warmup(self):
+        sa = self._resolve(enable_breakable_cuda_graph=True)
         self.assertEqual(sa.warmup_mode, "server")
 
     def test_disagg_role_disables_server_warmup(self):
@@ -659,10 +591,7 @@ class TestWarmupModeNormalization(unittest.TestCase):
         sa = self._resolve(
             warmup_mode="server",
             disagg_role=RoleType.DENOISER,
-            explicit=("warmup_mode",),
         )
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
         self.assertEqual(sa.warmup_mode, "request")
 
     def test_torch_compile_server_warmup_disabled_for_disagg_role(self):
@@ -670,12 +599,10 @@ class TestWarmupModeNormalization(unittest.TestCase):
 
         sa = self._resolve(enable_torch_compile=True, disagg_role=RoleType.DENOISER)
         self.assertEqual(sa.warmup_mode, "request")
-        self.assertTrue(sa.warmup)
-        self.assertFalse(sa.server_warmup)
 
     def test_invalid_mode_raises(self):
         with self.assertRaises(ValueError):
-            self._resolve(warmup_mode="bogus", explicit=("warmup_mode",))
+            self._resolve(warmup_mode="bogus")
 
 
 class TestWarmupImageIsModelValid(unittest.TestCase):
@@ -1962,24 +1889,9 @@ class TestPerRoleParallelism(unittest.TestCase):
         self.assertIsNone(par["ulysses_degree"])
         self.assertIsNone(par["ring_degree"])
 
-    def test_decoder_tp_is_alias_of_decoder_sp(self):
-        args = self._from_dict({"model_path": "/fake", "decoder_tp": 2})
-        from sglang.multimodal_gen.runtime.disaggregation.roles import RoleType
-
-        self.assertEqual(args.decoder_sp, 2)
-        par = args.get_role_parallelism(RoleType.DECODER)
-        self.assertIsNone(par["tp_size"])
-        self.assertEqual(par["sp_degree"], 2)
-
-    def test_conflicting_decoder_tp_and_decoder_sp_raise(self):
-        with self.assertRaisesRegex(ValueError, "decoder_tp is deprecated"):
-            self._from_dict(
-                {
-                    "model_path": "/fake",
-                    "decoder_tp": 2,
-                    "decoder_sp": 4,
-                }
-            )
+    def test_removed_decoder_tp_is_rejected(self):
+        with self.assertRaisesRegex(ValueError, "decoder_tp.*decoder_sp"):
+            self._from_dict({"model_path": "/fake", "decoder_tp": 2})
 
     def test_monolithic_returns_all_none(self):
         args = self._from_dict({"model_path": "/fake", "encoder_tp": 2})
@@ -2090,7 +2002,6 @@ class TestPerRoleParallelism(unittest.TestCase):
         self.assertEqual(args.denoiser_ring, 2)
         self.assertEqual(args.encoder_tp, 1)
         self.assertEqual(args.decoder_sp, 8)
-        self.assertIsNone(args.decoder_tp)
 
 
 class TestPipelineResolutionCliOverride(unittest.TestCase):