Record ASR smoke test success

docs/dependency-inventory.md

@@ -123,8 +123,8 @@ No `requirements.txt` found.
 
 | Model / asset | Source | Target location | Notes |
 | --- | --- | --- | --- |
-| Qwen ASR | `Qwen/Qwen3-ASR-1.7B` | Hugging Face/model cache or custom local model path | Loaded by `VoiceBridgeASRLoader`. |
-| Qwen forced aligner | `Qwen/Qwen3-ForcedAligner-0.6B` | Hugging Face/model cache or custom local model path | Loaded by `VoiceBridgeASRLoader`. |
+| Qwen ASR | `Qwen/Qwen3-ASR-1.7B` | persistent target `/data/voicegate_models/Qwen3-ASR/Qwen3-ASR-1.7B`; ComfyUI link root `ComfyUI/models/Qwen3-ASR` | Loaded by `VoiceBridgeASRLoader`. |
+| Qwen forced aligner | `Qwen/Qwen3-ForcedAligner-0.6B` | persistent target `/data/voicegate_models/Qwen3-ASR/Qwen3-ForcedAligner-0.6B`; ComfyUI link root `ComfyUI/models/Qwen3-ASR` | Loaded by `VoiceBridgeASRLoader`. |
 | VoxCPM2 | `openbmb/VoxCPM2` | persistent target `/data/voicegate_models/voxcpm/VoxCPM2`; ComfyUI link `ComfyUI/models/voxcpm/VoxCPM2` | RunningHub VoxCPM README documents the ComfyUI location. Approx. 4.6 GB. |
 | MelBand RoFormer | `Kijai/MelBandRoFormer_comfy` | persistent target `/data/voicegate_models/diffusion_models/MelBandRoFormer_comfy`; ComfyUI link `ComfyUI/models/diffusion_models/MelBandRoFormer_comfy` | Workflow references `MelBandRoFormer_comfy/MelBandRoformer_fp32.safetensors`. |
 
@@ -196,7 +196,10 @@ ComfyUI/
 |   `-- ComfyUI_AudioTools/
 `-- models/
     |-- voxcpm/VoxCPM2/
-    `-- diffusion_models/MelBandRoFormer_comfy/
+    |-- diffusion_models/MelBandRoFormer_comfy/
+    `-- Qwen3-ASR/
+        |-- Qwen3-ASR-1.7B/
+        `-- Qwen3-ForcedAligner-0.6B/
 ```
 
 Model files should be downloaded at runtime or first startup with

docs/hf-space-comfyui-step-by-step.zh.md

@@ -131,6 +131,9 @@ python scripts/bootstrap_comfy.py --with-models
 ```text
 ComfyUI/models/voxcpm/VoxCPM2
   -> /data/project_models/voxcpm/VoxCPM2
+
+ComfyUI/models/Qwen3-ASR
+  -> /data/project_models/Qwen3-ASR
 ```
 
 ## 6. 写最小 Gradio 入口
@@ -345,6 +348,27 @@ Gradio 进程和后续子进程可能看到不同 torch 版本。
 - 尽量复用模型进程。
 - 先做子流程验证，再优化 UX。
 
+### ZeroGPU duration 和配额预检
+
+`@spaces.GPU(duration=...)` 不要随手写得太大。VoiceGate 的 ASR 测试曾经
+使用 `duration=1800`，被 Space 直接拒绝：
+
+```text
+ZeroGPU illegal duration
+```
+
+后来改成 `duration=1200`，又因为剩余额度不足被配额预检拒绝：
+
+```text
+You have exceeded your Pro ZeroGPU quota
+```
+
+最终 ASR 诊断用 `duration=900` 跑通。建议做法：
+
+- 先用短输入和较小 duration 验证链路。
+- 模型下载放到非 GPU 的 Prepare 阶段。
+- 只有确认真实执行时间后，再提高 duration。
+
 ## 13. 最终正式化建议
 
 诊断阶段可以把很多按钮放在 `app.py` 里。正式阶段建议拆分：

docs/todo.md

@@ -109,7 +109,7 @@ from scaffold to a working Gradio app.
 - [x] Run a MelBand RoFormer smoke test.
 - [x] Run a MelBand RoFormer smoke test inside ZeroGPU.
 - [x] Run a tiny TTS-only workflow inside ZeroGPU.
-- [ ] Run a short ASR-only workflow.
+- [x] Run a short ASR-only workflow inside ZeroGPU.
 - [ ] Run SRT split -> VoxCPM -> SRT merge.
 - [ ] Run the full short-audio VoiceGate workflow.
 - [ ] Confirm output audio and SRT files are downloadable from Gradio.
@@ -127,6 +127,8 @@ from scaffold to a working Gradio app.
 - [x] Which exact node repository provides `RH_LLMAPI_NODE`?
 - [x] Which exact node repository provides `RunningHub_VoxCPM_*`?
 - [ ] Is `flash_attention_2` available and reliable in the ZeroGPU environment?
+- [x] Can ASR run without `flash_attention_2`? Yes. The ASR-only smoke test
+  used `attention=sdpa`.
 - [ ] Does VoxCPM2 fit comfortably in ZeroGPU memory with ASR and
   MelBandRoFormer in the same run?
 - [x] Where should large model files live? `/data/voicegate_models`, with

docs/work-log.md

@@ -12,7 +12,7 @@ and how they were resolved.
 - Space SDK: Gradio
 - Local Space wrapper repo: `VoiceGate-hf`
 - Local upstream reference checkout: `VoiceGate/`
-- Latest confirmed normal runtime commit: `6e4cd3f53d33e87979bdc814e0f1c34e0860f562`
+- Latest confirmed normal runtime commit: `316b35db739d74d05543d6c8c9dd9c16e0580b17`
 - Current expected Space secret: `DEEPSEEK_API_KEY`
 - Default persistent model root: `/data/voicegate_models`
 
@@ -35,12 +35,12 @@ Confirmed working:
 - DeepSeek-compatible LLM node with the Space secret.
 - MelBand RoFormer smoke tests in CPU mode and ZeroGPU mode.
 - VoxCPM2 TTS-only smoke test in ZeroGPU mode.
-- Persistent model storage for VoxCPM2 and MelBand under `/data`.
+- VoiceBridge ASR-only smoke test in ZeroGPU mode.
+- Persistent model storage for VoxCPM2, MelBand, Qwen3-ASR, and Qwen3 forced
+  aligner under `/data`.
 
 Not yet confirmed at the start of 2026-06-06:
 
-- VoiceBridge ASR-only workflow.
-- Qwen ASR and Qwen forced aligner model download/cache behavior in the Space.
 - SRT split -> VoxCPM -> SRT merge.
 - Full short-audio VoiceGate workflow.
 - Final user-facing Gradio upload/download UI.
@@ -265,6 +265,37 @@ This is workable for diagnostics, but final production should avoid heavy
 runtime `pip install` where possible. Prefer moving stable dependencies into
 Space build-time requirements or explicitly controlling pins.
 
+### ZeroGPU Duration and Quota Pitfall
+
+The ASR diagnostic was first decorated with:
+
+```python
+@spaces.GPU(duration=1800)
+```
+
+The Space rejected it before execution:
+
+```text
+ZeroGPU illegal duration
+The requested GPU duration is larger than the maximum allowed
+```
+
+After reducing the function to `duration=1200`, the Space still rejected the
+call because the quota precheck reported:
+
+```text
+You have exceeded your Pro ZeroGPU quota (1800s requested vs. 1389s left)
+```
+
+The working diagnostic used:
+
+```python
+@spaces.GPU(duration=900)
+```
+
+For future tests, keep diagnostic durations conservative and increase only when
+the workflow has already proven it needs more time.
+
 ## Dependency Pitfalls and Fixes
 
 `ComfyUI_AudioTools` initially failed to import.
@@ -558,6 +589,71 @@ run was very slow. The final app should minimize cold starts, avoid repeated
 ComfyUI/model reloads where possible, and use shorter diagnostic prompts while
 tuning.
 
+### VoiceBridge ASR-only
+
+On 2026-06-06, a minimal VoiceBridge ASR-only workflow was run from the normal
+Gradio runtime inside `@spaces.GPU`.
+
+Before running ASR, `scripts/bootstrap_comfy.py` was extended so Qwen ASR models
+also live on persistent storage:
+
+```text
+/home/user/app/ComfyUI/models/Qwen3-ASR
+  -> /data/voicegate_models/Qwen3-ASR
+```
+
+The model preparation downloads:
+
+```text
+/data/voicegate_models/Qwen3-ASR/Qwen3-ASR-1.7B
+/data/voicegate_models/Qwen3-ASR/Qwen3-ForcedAligner-0.6B
+```
+
+Minimal workflow:
+
+```text
+LoadAudio
+  -> VoiceBridgeASRLoader(attention=sdpa, forced_aligner=Qwen/Qwen3-ForcedAligner-0.6B)
+  -> VoiceBridgeASRTranscribe(return_timestamps=True)
+  -> GenerateSRT
+  -> easy showAnything
+```
+
+Input:
+
+```text
+D:\voicebridge-test-audio\test_audio\2-坤哥.MP3
+```
+
+Result:
+
+```text
+status_str=success
+completed=True
+elapsed_sec=62.4
+```
+
+Returned SRT text:
+
+```text
+1
+00:00:02,080 --> 00:00:03,200
+全民制作人们 大家好
+
+2
+00:00:03,439 --> 00:00:06,160
+我是练习时长两年半的个人练习生蔡徐坤
+
+3
+00:00:06,480 --> 00:00:09,359
+喜欢唱、跳、rap、篮球、music
+```
+
+This confirms the Qwen3-ASR model, forced aligner, VoiceBridge ASR nodes, and
+SRT generation can run in the Space ZeroGPU path. The smoke test intentionally
+used `attention=sdpa` instead of `flash_attention_2`; `flash_attention_2`
+availability remains unverified.
+
 ## Secrets and API Keys
 
 `DEEPSEEK_API_KEY` should be stored only as a Hugging Face Space Secret.
@@ -609,6 +705,9 @@ ComfyUI/models/voxcpm/VoxCPM2
 
 ComfyUI/models/diffusion_models/MelBandRoFormer_comfy
   -> /data/voicegate_models/diffusion_models/MelBandRoFormer_comfy
+
+ComfyUI/models/Qwen3-ASR
+  -> /data/voicegate_models/Qwen3-ASR
 ```
 
 Override the root with:
@@ -624,6 +723,8 @@ persistent storage:
 /data/voicegate_models/voxcpm/VoxCPM2/model.safetensors
 /data/voicegate_models/voxcpm/VoxCPM2/audiovae.pth
 /data/voicegate_models/diffusion_models/MelBandRoFormer_comfy/MelBandRoformer_fp32.safetensors
+/data/voicegate_models/Qwen3-ASR/Qwen3-ASR-1.7B
+/data/voicegate_models/Qwen3-ASR/Qwen3-ForcedAligner-0.6B
 ```
 
 Verified symlinks:
@@ -634,6 +735,9 @@ Verified symlinks:
 
 /home/user/app/ComfyUI/models/diffusion_models/MelBandRoFormer_comfy
   -> /data/voicegate_models/diffusion_models/MelBandRoFormer_comfy
+
+/home/user/app/ComfyUI/models/Qwen3-ASR
+  -> /data/voicegate_models/Qwen3-ASR
 ```
 
 `DEEPSEEK_API_KEY` was also verified as present in the Space environment without
@@ -655,23 +759,15 @@ Model download pitfall:
 - `d849d03` Record ComfyUI API smoke test
 - `79b8b37` Add GPU smoke test button
 - `6e4cd3f` Run Space preparation in background
+- `b39ef30` Add ASR diagnostic workflow and deployment guide
+- `316b35d` Reduce ASR ZeroGPU duration
 
 ## Remaining Work
 
 Next recommended steps:
 
-1. Verify the `DEEPSEEK_API_KEY` secret is visible to the Space process without
-   printing its value.
-2. Add model download or model presence checks for:
-   - VoxCPM2
-   - MelBand RoFormer
-   - Qwen ASR
-   - Qwen forced aligner
-3. Wire `app.py` to bootstrap and start ComfyUI from the Gradio process.
-4. Wrap the actual VoiceGate execution path in `@spaces.GPU`.
-5. Run progressively larger workflows:
-   - minimal API workflow
-   - TTS-only
-   - ASR-only
+1. Run progressively larger workflows:
    - SRT split and merge
    - full short-audio VoiceGate workflow
+2. Replace the diagnostic panel with the first real Gradio upload/download UI.
+3. Reduce runtime dependency installation and model reload overhead.