[New Model] Add support for tencent/Covo-Audio-Chat by Dnoob · Pull Request #2293 · vllm-project/vllm-omni

Dnoob · 2026-03-28T05:27:18Z

The model to consider

Model Weights: https://huggingface.co/tencent/Covo-Audio-Chat
Model Code: https://github.com/Tencent/Covo-Audio

Model description

This PR adds support for Covo-Audio-Chat (Tencent, 7B end-to-end audio language model) with a 2-stage pipeline:

Stage 0 (fused_thinker_talker)

Whisper encoder + AudioAdapter + Qwen2.5-7B LLM → interleaved text + audio tokens

Stage 1 (code2wav)

BigVGAN vocoder → 24kHz audio waveform

Note: This model requires audio input to produce correctly interleaved text + audio tokens. Text-only input causes abnormal interleaving patterns, resulting in incomplete audio output. This is a model design constraint, not a code limitation.

Changes

Core model: covo_audio.py (dual-stage router), covo_audio_llm.py (Stage 0), covo_audio_code2wav.py (Stage 1)
Config: config_covo_audio.py, stage YAML configs
Prompt utilities: prompt_utils.py (centralized prompt templates and construction helpers)
Stage input processor: extracts audio codes from Stage 0 output for Stage 1
Registry + speech endpoint integration
token2wav.py (consolidated from upstream model repo into a single module, import paths adapted)
Speaker prompt .npy files
Offline inference example + OpenAI-compatible client example
Documentation update to supported_models.md

Fixes #2004

Test plan

E2E test (requires GPU with ≥20 GiB VRAM):

CUDA_VISIBLE_DEVICES=0 pytest -s -v tests/e2e/offline_inference/test_covo_audio.py::test_audio_to_audio

Online serving:

CUDA_VISIBLE_DEVICES=0 vllm serve tencent/Covo-Audio-Chat --omni \
    --stage-configs-path vllm_omni/model_executor/stage_configs/covo_audio.yaml \
    --trust-remote-code

python examples/online_serving/covo_audio/openai_chat_completion_client.py

Test result

tests/e2e/offline_inference/test_covo_audio.py::test_audio_to_audio[omni_runner0] PASSED
======================= 1 passed in 62.86s ========================

Environment

GPU: 1x A100 (80 GiB)
Stage 0 (7B LLM): ~16 GiB VRAM
Stage 1 (BigVGAN): ~2 GiB VRAM

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5b064383be

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".

Dnoob · 2026-03-28T05:44:54Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cf389cbb49

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 · 2026-03-28T05:51:12Z

+    engine_args:
+      model_stage: fused_thinker_talker
+      max_num_seqs: 1
+      model_arch: CovoAudioForConditionalGeneration


Use a registered Covo model_arch in stage config

The new stage config sets model_arch: CovoAudioForConditionalGeneration, but this architecture key is not registered in vllm_omni/model_executor/models/registry.py (the commit only registers CovoAudioForCausalLM and CovoAudioModel as keys that map to the Covo class). Because OmniEngineArgs.create_model_config() injects model_arch directly into architectures, serving with vllm_omni/model_executor/stage_configs/covo_audio.yaml can fail model resolution at startup instead of loading Covo; the config should use a registered key or the registry should add CovoAudioForConditionalGeneration explicitly.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-03-28T05:51:12Z

 _QWEN3_TTS_MODEL_STAGES = {"qwen3_tts"}
 _FISH_TTS_MODEL_STAGES = {"fish_speech_slow_ar"}
-_TTS_MODEL_STAGES: set[str] = _VOXTRAL_TTS_MODEL_STAGES | _QWEN3_TTS_MODEL_STAGES | _FISH_TTS_MODEL_STAGES
+_COVO_AUDIO_MODEL_STAGES = {"fused_thinker_talker"}


Restrict Covo TTS stage matching to Covo models

Marking fused_thinker_talker as a Covo TTS stage is too broad: the same model_stage is used by non-Covo pipelines (for example, MiMoAudio stage configs), so _find_tts_stage/_detect_tts_model_type will classify those models as covo_audio and route /v1/audio/speech through Covo-specific prompt construction. This can produce incorrect prompts or runtime failures for other models that share the stage name; model detection should include an architecture/model check instead of relying on stage name alone.

Useful? React with 👍 / 👎.

linyueqian

Tested locally on A100-80G. A few findings:

Bug: model path resolution breaks with HF repo names

covo_audio_code2wav.py:26 uses vllm_config.model_config.model directly as a filesystem path. When the model is specified as a HF repo name (e.g. tencent/Covo-Audio-Chat), os.path.join(model_path, "token2wav", ...) resolves to a relative path that doesn't exist:

FileNotFoundError: [Errno 2] No such file or directory: 'tencent/Covo-Audio-Chat/token2wav/global_mean_var.npy'

Need to resolve to the local cache path first, e.g.:

if os.path.isdir(model_name):
    model_path = model_name
else:
    from huggingface_hub import snapshot_download
    model_path = snapshot_download(model_name)

Missing dependency: torchdiffeq

The vendored token2wav flow matching module requires torchdiffeq, but it's not declared in pyproject.toml. Stage 1 fails with ModuleNotFoundError: No module named 'torchdiffeq'.

Audio output cuts off abruptly

With the CI config (max_tokens: 512, ignore_eos: true), the text-to-audio test produces ~9.25s of audio that ends mid-sentence. The interleaving ratio (5 text + 15 audio tokens) means only ~384 of the 512 tokens are audio codes. Consider either increasing max_tokens or removing ignore_eos: true so the model can stop naturally.

Minor

PR is 42 commits behind main, needs rebase
test_audio_to_audio requires espeak-ng system package (via pyttsx3 for synthetic audio generation). Worth noting in the test plan or adding a skip condition.

test_text_to_audio passes after fixing the path issue. Model produces reasonable text + audio output.

Dnoob · 2026-03-30T06:18:48Z

Tested locally on A100-80G. A few findings:

Bug: model path resolution breaks with HF repo names

covo_audio_code2wav.py:26 uses vllm_config.model_config.model directly as a filesystem path. When the model is specified as a HF repo name (e.g. tencent/Covo-Audio-Chat), os.path.join(model_path, "token2wav", ...) resolves to a relative path that doesn't exist:
FileNotFoundError: [Errno 2] No such file or directory: 'tencent/Covo-Audio-Chat/token2wav/global_mean_var.npy'
Need to resolve to the local cache path first, e.g.:
if os.path.isdir(model_name):
    model_path = model_name
else:
    from huggingface_hub import snapshot_download
    model_path = snapshot_download(model_name)
Missing dependency: torchdiffeq

The vendored token2wav flow matching module requires torchdiffeq, but it's not declared in pyproject.toml. Stage 1 fails with ModuleNotFoundError: No module named 'torchdiffeq'.

Audio output cuts off abruptly

With the CI config (max_tokens: 512, ignore_eos: true), the text-to-audio test produces ~9.25s of audio that ends mid-sentence. The interleaving ratio (5 text + 15 audio tokens) means only ~384 of the 512 tokens are audio codes. Consider either increasing max_tokens or removing ignore_eos: true so the model can stop naturally.

Minor

PR is 42 commits behind main, needs rebase

test_audio_to_audio requires espeak-ng system package (via pyttsx3 for synthetic audio generation). Worth noting in the test plan or adding a skip condition.

test_text_to_audio passes after fixing the path issue. Model produces reasonable text + audio output.

For the torchdiffeq dependency, should it go into requirements/common.txt (which affects all users), or would you prefer a lazy import with a clear error message + doc note, since it's only used by the vendored token2wav module?

linyueqian · 2026-04-02T00:31:29Z

@amy-why-3459 PTAL

linyueqian · 2026-04-02T00:33:40Z

Thanks for the fixes on path resolution and max_tokens.

For torchdiffeq — go with a lazy import since it's only used by the vendored token2wav:

try:
    from torchdiffeq import odeint
except ImportError:
    raise ImportError(
        "Covo-Audio code2wav requires `torchdiffeq`. "
        "Install it with: pip install torchdiffeq"
    )

Still outstanding:

Rebase onto main (currently blocked)
espeak-ng — add a skip condition for test_audio_to_audio

amy-why-3459 · 2026-04-02T01:07:28Z

Thank you very much for your contribution. Could you please add a readme file for the model?

amy-why-3459 · 2026-04-02T01:10:54Z

+from .token2latent import Token2latentFlowMatchingWithEmbed
+
+
+class Token2WavDecoder(BaseModel):


Can we put these function definitions in the model file to avoid creating too many folders?

Sorry, I didn't consider this earlier and directly copied the structure from the model repo, which made it bloated. I've analyzed it and plan to consolidate the entire token2wav/ folder into a single token2wav.py file and remove unnecessary code. Will update in the next push.

hsliuustc0106

it seems there are a lot of dead code inside this PR

Dnoob · 2026-04-02T13:55:08Z

Thank you very much for your contribution. Could you please add a readme file for the model?

Already added

lishunyang12

left a few comments on the core model files. the vendored token2wav code is fine to skip lint-wise but the non-vendored parts have a couple issues.

lishunyang12 · 2026-04-02T15:37:58Z

+        inputs_embeds: torch.Tensor | None = None,
+        generate_audio: bool = True,
+        codec: torch.Tensor | None = None,
+        sampling_metadata: SamplingMetadata | None = None,


This materializes the entire safetensors weight iterator into a Python list just to filter by prefix. For a 7B model that's ~14 GB of extra peak memory. Use a generator instead:

Suggested change

sampling_metadata: SamplingMetadata | None = None,

llm_weights = ((k, v) for k, v in weights if k.startswith(("llm", "encoder", "audio_adapter")))

lishunyang12 · 2026-04-02T15:37:58Z

+                **kwargs,
+            )
+
+            return OmniOutput(


Several of these forward params (generate_audio, codec, logits_index, sampler, additional_information) are never used. Remove them — dead parameters in the forward signature are confusing for anyone reading the dispatch logic.

lishunyang12 · 2026-04-02T15:37:58Z

+        logger.info(
+            "Request %s: total_tokens=%d, text_tokens=%d, audio_tokens=%d",
+            request_id,
+            len(token_ids),


This is a lot of per-request logging at INFO level (token counts, full interleaving pattern). In production with concurrent requests this will flood the logs. Drop the pattern log to DEBUG, or remove it.

Adapt Covo-Audio-Chat (Tencent, 7B end-to-end audio language model) to vllm-omni with a 2-stage pipeline: - Stage 0 (fused_thinker_talker): Whisper encoder + AudioAdapter + Qwen2.5-7B LLM → interleaved text + audio tokens - Stage 1 (code2wav): BigVGAN vocoder → 24kHz audio waveform Closes vllm-project#2004 Signed-off-by: Dnoob <dxpouo@gmail.com>

Signed-off-by: Dnoob <dxpouo@gmail.com>

… for Covo-Audio Signed-off-by: Dnoob <dxpouo@gmail.com>

…Audio-Chat - Add prompt_utils.py with shared system prompt and prompt builders, removing 3 duplicates - Add offline inference example (end2end.py + README) - Fix online client: add system prompt, stop_token_ids, ignore_eos, detokenize=false - Fix stage config detokenize setting for code2wav - Use local sample audio for online example instead of S3 download - Add --port 18091 to online README to match client config Signed-off-by: Dnoob <dxpouo@gmail.com>

Signed-off-by: Dnoob <dxpouo@gmail.com>

- Use generator instead of list comprehension in load_weights to reduce peak memory - Remove unused forward parameters (generate_audio, codec, logits_index, sampler, additional_information) - Remove debug logging from stage input processor Signed-off-by: Dnoob <dxpouo@gmail.com>

- Fix MultiModalDataDict import path for vllm 0.19.0 compatibility - Remove unsupported text-only test (model requires audio input) - Use sample_audio.wav instead of pyttsx3 synthetic audio in test - Remove espeak-ng dependency from test Signed-off-by: Dnoob <dxpouo@gmail.com>

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob · 2026-04-04T10:55:02Z

@lishunyang12 @amy-why-3459 Addressed all review feedback, rebased on main, and updated PR description. PTAL, thanks!

linyueqian

LGTM

yenuo26 · 2026-04-07T09:23:56Z

+    return "请回答这段音频里的问题。"
+
+
+@pytest.mark.core_model


if this want to be added to nightly, please:
1.change label to @pytest.mark.advanced_model
2.rename this script to test_covo_audio_expansion.py

yenuo26 · 2026-04-07T09:26:17Z

+        }
+    ]
+
+    outputs = omni_runner.generate(omni_inputs)


maybe you can use omni_runner_handler.send_request(request_config) in conftest.py, like tests/e2e/offline_inference/test_qwen3_omni.py

yenuo26 · 2026-04-07T09:28:47Z

+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/e2e/offline_inference/test_covo_audio.py -m "core_model" --run-level "core_model"


i think you can modify commands in 🌕 Omni Model Test with H100, instead of adding a separate new job.

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob · 2026-04-08T04:44:51Z

@yenuo26 Sorry, I forgot to add the test file under online_serving/.

I checked how other omni models organize their tests in this repo. The convention seems to be: offline advanced_model tests go in test-merge.yml, and online _expansion.py tests go in test-nightly.yml.

Adjusted accordingly:

Added tests/e2e/online_serving/test_covo_audio_expansion.py
Removed the separate Covo-Audio-Chat job and reused the existing "Omni Model Test with H100"

Signed-off-by: Dnoob <dxpouo@gmail.com>

linyueqian · 2026-04-13T15:29:07Z

fix ci please

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob · 2026-04-15T01:00:56Z

@linyueqian @hsliuustc0106 CI failures are fixed and all checks pass. PTAL, thanks!

hsliuustc0106 · 2026-04-17T12:32:09Z

 | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | |
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | |
 | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | |
+| `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | |


nit: This PR includes online serving support (OpenAI-compatible client example + test_covo_audio_expansion.py), so the Online column should be ✅︎ instead of empty, to match the other models in this table.

I don't see an "Online" column in this table, just NVIDIA GPU、AMD GPU、Ascend NPU、Intel GPU. Which one did you mean?

hsliuustc0106 · 2026-04-17T12:34:41Z

+        self.o_dropout = Dropout(dropout)
+
+        self.cpu_config = AttentionConfig(True, True, True)
+        device_properties = torch.cuda.get_device_properties(torch.device("cuda"))


nit (vendored code): torch.cuda.get_device_properties(torch.device("cuda")) hardcodes device index 0. In a multi-GPU setup where this vocoder runs on a non-default CUDA device, this may select the wrong flash-attention config. I understand this is vendored from upstream, so just flagging for awareness — no action required for this PR.

hsliuustc0106

Blocker Scan

Category	Result
Correctness	PASS
Reliability/Safety	PASS
Breaking Changes	PASS
Test Coverage	PASS (offline test in PR desc, CI green, nightly test added)
Documentation	PASS (supported_models.md, offline + online examples + README)
Security	PASS

Merge Gate

DCO: PASS
pre-commit: PASS
Build: PASS
Buildkite CI (amd/intel): PASS
Mergeable: FAIL — CONFLICTING, needs rebase onto latest main

Summary

The code is well-structured after multiple review rounds. All previously flagged issues have been addressed:

HF repo path resolution fixed (snapshot_download fallback)
torchdiffeq lazy import with clear error message
Dead parameters removed from forward signatures
INFO-level log flooding fixed
llm_weights uses generator instead of list materialization
Vendored token2wav/ consolidated into single token2wav.py
CovoAudioForConditionalGeneration registered in model registry
CI test config updated with max_tokens: 2048

Left 2 nits as inline comments (supported_models table + vendored code awareness).

The only remaining blocker is the merge conflict. Please rebase onto latest main and I'll approve.

# Conflicts: # vllm_omni/entrypoints/openai/serving_speech.py

Dnoob · 2026-04-19T12:21:00Z

@hsliuustc0106 PTAL, thanks!

hsliuustc0106 · 2026-04-19T12:24:22Z

 [tool.ruff.lint.per-file-ignores]
 "examples/**" = ["E501"]  # Allow long lines in examples
 "tests/**" = ["E501"]  # Allow long lines in tests
+"**/token2wav/**" = ["E501", "E721", "E741", "F401", "F403", "F405", "F841", "UP028"]  # Vendored third-party code


why add this line?

hsliuustc0106 · 2026-04-19T12:25:15Z

@@ -0,0 +1,78 @@
+#


please follow the new pipeline with #2383 merged

hsliuustc0106 · 2026-04-19T12:26:20Z

also, please add a new recipe

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob · 2026-04-21T03:48:40Z

@hsliuustc0106 All addressed and CI is green. PTAL, thanks!

lishunyang12 · 2026-04-23T06:42:10Z

Resolve conflict.

Signed-off-by: Dnoob <dxpouo@gmail.com>

…itespace Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob requested a review from hsliuustc0106 as a code owner March 28, 2026 05:27

Dnoob force-pushed the feat/covo-audio-chat branch 2 times, most recently from a142631 to e44e30e Compare March 28, 2026 05:35

chatgpt-codex-connector Bot reviewed Mar 28, 2026

View reviewed changes

Comment thread vllm_omni/model_executor/models/covo_audio/covo_audio_llm.py

Comment thread vllm_omni/model_executor/models/covo_audio/covo_audio.py Outdated

Dnoob force-pushed the feat/covo-audio-chat branch from e44e30e to cf389cb Compare March 28, 2026 05:43

chatgpt-codex-connector Bot reviewed Mar 28, 2026

View reviewed changes

Dnoob force-pushed the feat/covo-audio-chat branch 2 times, most recently from 1232884 to b372618 Compare March 28, 2026 06:12

linyueqian reviewed Mar 30, 2026

View reviewed changes

amy-why-3459 reviewed Apr 2, 2026

View reviewed changes

Dnoob force-pushed the feat/covo-audio-chat branch from 8840f97 to 7d6696c Compare April 2, 2026 05:48

hsliuustc0106 reviewed Apr 2, 2026

View reviewed changes

lishunyang12 reviewed Apr 2, 2026

View reviewed changes

Dnoob added 6 commits April 4, 2026 09:31

fix model path resolution and increase CI max_tokens for Covo-Audio-Chat

3c25d66

Signed-off-by: Dnoob <dxpouo@gmail.com>

[Bugfix] Add lazy import for torchdiffeq and espeak-ng skip condition…

f1b1b3f

… for Covo-Audio Signed-off-by: Dnoob <dxpouo@gmail.com>

[Refactor] Consolidate token2wav package into single module

bbb8ef6

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob force-pushed the feat/covo-audio-chat branch from b0f956b to 9600536 Compare April 4, 2026 09:32

Dnoob added 2 commits April 4, 2026 10:33

[Style] Fix ruff lint and formatting issues

56289a3

Signed-off-by: Dnoob <dxpouo@gmail.com>

linyueqian approved these changes Apr 5, 2026

View reviewed changes

yenuo26 reviewed Apr 7, 2026

View reviewed changes

add online expansion test and consolidate nightly job

bed6057

Signed-off-by: Dnoob <dxpouo@gmail.com>

Merge branch 'main' into feat/covo-audio-chat

5b79c4f

hsliuustc0106 enabled auto-merge (squash) April 9, 2026 07:02

Merge branch 'main' into feat/covo-audio-chat

2a3169e

Signed-off-by: Dnoob <dxpouo@gmail.com>

auto-merge was automatically disabled April 13, 2026 07:12
Head branch was pushed to by a user without write access

fix: resolve merge conflict syntax error in registry.py

89c004b

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob added 2 commits April 14, 2026 02:33

[CI] Fix pytest collection failure caused by covo_audio test

2978892

Signed-off-by: Dnoob <dxpouo@gmail.com>

Merge branch 'main' into feat/covo-audio-chat

ea21213

Signed-off-by: Dnoob <dxpouo@gmail.com>

hsliuustc0106 reviewed Apr 17, 2026

View reviewed changes

Merge remote-tracking branch 'origin/main' into feat/covo-audio-chat

f8ba351

# Conflicts: # vllm_omni/entrypoints/openai/serving_speech.py

hsliuustc0106 reviewed Apr 19, 2026

View reviewed changes

Dnoob added 6 commits April 20, 2026 06:10

remove token2wav ruff ignore rule

35dc318

Signed-off-by: Dnoob <dxpouo@gmail.com>

Merge remote-tracking branch 'origin/main' into feat/covo-audio-chat

b68220a

migrate covo_audio to new pipeline/deploy config schema

8f35f14

Signed-off-by: Dnoob <dxpouo@gmail.com>

add recipe for covo_audio

d78b6ff

Signed-off-by: Dnoob <dxpouo@gmail.com>

Merge branch 'main' into feat/covo-audio-chat

e15cef1

Signed-off-by: Dnoob <dxpouo@gmail.com>

fix ci import errors

e018aec

Signed-off-by: Dnoob <dxpouo@gmail.com>

Dnoob added 3 commits April 23, 2026 15:51

Merge branch 'main' into feat/covo-audio-chat

584034c

Signed-off-by: Dnoob <dxpouo@gmail.com>

Merge branch 'main' into feat/covo-audio-chat

2e86a7d

fix: replace librosa with vllm.multimodal helper and drop trailing wh…

275bde7

…itespace Signed-off-by: Dnoob <dxpouo@gmail.com>

		from .token2latent import Token2latentFlowMatchingWithEmbed


		class Token2WavDecoder(BaseModel):

	sampling_metadata: SamplingMetadata \| None = None,
	llm_weights = ((k, v) for k, v in weights if k.startswith(("llm", "encoder", "audio_adapter")))

		return "请回答这段音频里的问题。"


		@pytest.mark.core_model

Conversation

Dnoob commented Mar 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

The model to consider

Model description

Stage 0 (fused_thinker_talker)

Stage 1 (code2wav)

Changes

Test plan

Test result

Environment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Dnoob commented Mar 28, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Mar 28, 2026

Choose a reason for hiding this comment

Uh oh!

linyueqian left a comment

Choose a reason for hiding this comment

Uh oh!

Dnoob commented Mar 30, 2026

Uh oh!

linyueqian commented Apr 2, 2026

Uh oh!

linyueqian commented Apr 2, 2026

Uh oh!

amy-why-3459 commented Apr 2, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 left a comment

Choose a reason for hiding this comment

Uh oh!

Dnoob commented Apr 2, 2026

Uh oh!

lishunyang12 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Dnoob commented Apr 4, 2026

Uh oh!

linyueqian left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Dnoob commented Apr 8, 2026

Uh oh!

linyueqian commented Apr 13, 2026

Uh oh!

Dnoob commented Apr 15, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Dnoob commented Mar 28, 2026 •

edited

Loading