[feature] : add cache-dit for stable-audio-open-1.0

[akshatvishu]

Part of #1217

Purpose

Add cache-dit support for stable audio open 1.0

Test Plan

    omni = Omni(
        model=MODEL_PATH,
        dtype="float16",
        num_workers=1,
        cache_backend=cache_backend,
        cache_config=cache_config
    )
}

sampling_params = OmniDiffusionSamplingParams(
    num_inference_steps=100,
    guidance_scale=7.0,
    seed=42,
    extra_args={"audio_end_in_s": 10.0}
)

outputs = omni.generate(
    {"prompt": "The sound of a hammer hitting a wooden surface", "negative_prompt": "Low quality, noisy"},
    sampling_params
)

full comprehensive testing can be found in this kaggle_notebook

Test Result

• Device: cuda

• GPU: NVIDIA Tesla T4

• Prompt : The sound of a hammer hitting a wooden surface

• num_inference_steps=100

• guidance_scale=7.0,

• max_audio_length = 10 seconds

Baseline:

Configuration	Time	Speed Up(vs baseline)	file (mp3)
Baseline( OMNI)	25.91 s	-	baseline.mp3
Baseline (HF Diffuser )	27.78s	-	baseline_hf.mp3

Config1:

    "configA_balanced": {
        "Fn_compute_blocks": 2,
        "Bn_compute_blocks": 0,
        "max_warmup_steps": 4,
        "residual_diff_threshold": 0.22,
        "max_continuous_cached_steps": 3,
        "enable_taylorseer": True,
        "taylorseer_order": 1,
    },

Configuration	Time	Speed Up(vs baseline)	file (mp3)
OMNI	22.08s	1.17x	configA_balanced.mp3
HF Diffuser + CacheDit	24.69s	1.13x	configA_balanced_hf.mp3

Config2:

    "configB_aggressive": {
        "Fn_compute_blocks": 1,
        "Bn_compute_blocks": 0,
        "max_warmup_steps": 3,
        "residual_diff_threshold": 0.30,
        "max_continuous_cached_steps": 5,
        "enable_taylorseer": True,
        "taylorseer_order": 1,
    },

Configuration	Time	Speed Up(vs baseline)	file (mp3)
OMNI	20.15 s	1.29x	configB_aggressive.mp3
HF Diffuser + CacheDit	24.05s	1.16x	configB_aggressive_hf.mp3

Config3:

    "configC_ultra": {
        "Fn_compute_blocks": 1,
        "Bn_compute_blocks": 0,
        "max_warmup_steps": 2,
        "residual_diff_threshold": 0.35,
        "max_continuous_cached_steps": 6,
        "enable_taylorseer": True,
        "taylorseer_order": 2,
    }
}

Configuration	Time	Speed Up(vs baseline)	file (mp3)
OMNI	19.16 s	1.35x	configC_ultra.mp3
HF Diffuser + CacheDit	20.90s	1.33x	configC_ultra_hf.mp3

Files are in .mp3 format as github doesn't support .wav in comments.

Note :

• Stable Audio Open 1.0 exhibits a high natural step-to-step drift (median residual ≈0.34) as seen in cache-dit.summary() when running the same config as vllm-omni in hf diffuser+cache-dit setup. To achieve significant speedups on T4 hardware, it is necessary to use a residual_diff_threshold near or above this drift value as using conservative residual_diff_threshold like 0.12 resulted in 1.00x speedup (or even slowdowns) because the cache missed on nearly every step, leaving only the management overhead without any compute savings.

• The vllm-omni orchestrator performs a 1-step dummy warmup run during server initialization, If a user provides an SCM (Step Computation Masking) policy, the engine crashes with the following error:

AssertionError: Only total_steps=4 or 6 is supported for predefined masks while total_steps < 8. Got total_steps=1.

Thus, I am wondering if we should a guard condition like below or it's an acceptable behavior.

def refresh_cache_context(pipeline: Any, num_inference_steps: int, verbose: bool = True) -> None:
    """
    Refresh cache context. 
    Guards against 1-step dummy warmup causing SCM mask generation errors.
    """
    # Disable SCM policy for the 1-step dummy warmup to prevent AssertionError
    effective_mask_policy = cache_config.scm_steps_mask_policy if num_inference_steps > 1 else None
   

• Also added the missing _repeated_blocks = ["StableAudioDiTBlock"] to StableAudioDiTModel to enable regional compilation and backend patching.

---

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

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)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cf50517d5d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[hsliuustc0106]

fix DCO please

[akshatvishu]

@hsliuustc0106 Sorry! I've updated it !