[Bugfix] Fix Fish Speech startup on vLLM 0.18.0 - missing architectures

[linyueqian]

Summary

Fix Fish Speech (and other models without architectures in config.json) failing to start on vLLM 0.18.0.

vLLM 0.18.0 now requires architectures in the model config. Fish Speech s2-pro's config.json doesn't have this field, but the stage config already specifies model_arch. This fix auto-injects model_arch into hf_overrides before ModelConfig creation.

Test plan

# Before: crashes with "No model architectures are specified"
# After: starts successfully
python examples/offline_inference/fish_speech/end2end.py \
    --model fishaudio/s2-pro --text "Hello test"

Tested with s2-pro on A100, offline inference produces valid 1.02s audio @ 44.1kHz.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0fe68c8da6

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".

[chatgpt-codex-connector]

Avoid loading Hugging Face config twice during startup

This new pre-check calls get_config(...) before super().create_model_config(), and EngineArgs.create_model_config() will load the HF config again when constructing the vLLM ModelConfig. For stages that set model_arch and have no hf_overrides, this adds an extra remote/local config load per stage, which increases startup time and can introduce avoidable transient fetch failures for multi-stage pipelines. Prefer reusing a single loaded config or injecting/merging architectures without a second fetch path.

Useful? React with 👍 / 👎.

[linyueqian]

@Sy0307 is this patch necessary? strange that i need to add this for fish speech.

[gcanlin]

Please see #1898. Looks like we don't need it?

[Sy0307]

@Sy0307 is this patch necessary? strange that i need to add this for fish speech.

When I am checking out vllm 0.18.0, the same issue occured. This PR is necessary for me.

[gcanlin]

cc @princepride

[princepride]

I will check it later.

[princepride]

I believe this fix may not solve the real problem, because when we already give the model's architecture, there is no sense we need hf_override to override it.

[princepride]

The model_arch in the YAML does not take effect during vLLM's base ModelConfig.init phase.

Timing issue:

1. OmniEngineArgs picks up self.model_arch = "FishSpeechSlowARForConditionalGeneration" from the YAML.
2. It calls super().create_model_config() → vLLM's ModelConfig.init runs.
3. ModelConfig.init reads the architecture from hf_config.architectures → at this point model_arch has not been used yet.
4. If the HF config.json has no architectures field → crash.
5. OmniModelConfig.from_vllm_model_config(model_arch=...) is never reached.

model_arch only takes effect after step 5 (via the OmniModelConfig.architectures property override + the hf_config.architectures write-back), but because the fishaudio/s2-pro model's config.json is non-standard and lacks an architectures field, the program already crashes at step 4.

[princepride]

LGTM

[linyueqian]

Closing in favor of #2178 which was already merged with an equivalent fix.