[Recipe] zai-org/GLM-TTS (2x non-standard RTX 4090 48GB) (#5769)

Signed-off-by: harley <480497125@qq.com>
2026-08-26 22:16:49 +08:00
1 parent 8cfede4309
commit 926e7fbfa2
@@ -87,7 +87,7 @@ recipes/
 | [`Wan-AI/Wan2.1-VACE.md`](./Wan-AI/Wan2.1-VACE.md) | Unified T2V, I2V, V2LF, FLF2V, inpaint, and R2V | 1x RTX 5090 (1.3B) / 1x L40S 48GB with layerwise offload (14B) |
 | [`XiaomiMiMo/MiMo-Audio.md`](./XiaomiMiMo/MiMo-Audio.md) | Offline + online omni audio (TTS, ASR, dialogue) | 1x RTX 5090 / 5090D 32GB |
 | [`StabilityAI/Stable-Diffusion-3.5.md`](./StabilityAI/Stable-Diffusion-3.5.md) | Text-to-image serving (SD 3.5-medium and SD 3.5-large) | 1x RTX A6000 48GB |
-| [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB |
+| [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB / 2x non-standard RTX 4090 48GB (49,140 MiB each) |
 | [`GLM/GLM-Image.md`](./GLM/GLM-Image.md) | Online serving for image generation | 1x A800 80GB / 2x A800 80GB |
 | [`JD/JoyAI-VL-Interaction.md`](./JD/JoyAI-VL-Interaction.md) | Real-time streaming video-language interaction (proactive speak/silence/delegate) | 1x GPU 24GB+ |

@@ -1,4 +1,4 @@
-# GLM-TTS for Chinese/English TTS on 1x GPU
+# GLM-TTS for Chinese/English TTS

 ## Summary

@@ -103,3 +103,137 @@ python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
 - Key flags: `--omni` is required; `--trust-remote-code` is needed for the GLM-TTS phoneme tokenizer; the DiT stage enables bucketed model-internal CUDA graphs with eager fallback.
 - Voice cloning: requires `ref_audio` + `ref_text` together. Reference audio should be 3-10 seconds. Feature extraction (WhisperVQ tokenizer, CampPlus ONNX, mel) runs on the model side.
 - Known limitations: First request may be slow due to lazy model loading (WhisperVQ, CampPlus ONNX). Warm-cache RTF is approximately 0.6-0.7x on A40.
+
+### 2x non-standard RTX 4090 48GB
+
+This configuration was personally validated on two non-standard RTX 4090 cards
+reporting 49,140 MiB each in `nvidia-smi`. Stock 24GB RTX 4090 cards were not
+validated.
+
+#### Environment
+
+- OS: Ubuntu 22.04.1 LTS
+- NVIDIA driver: 580.76.05
+- Python: 3.12.13
+- PyTorch: 2.11.0+cu130
+- CUDA runtime and JIT toolkit: 13.0
+- vLLM: 0.26.0
+- vLLM-Omni: commit `66a9b0c8`, version
+  `0.26.0rc2.dev5+g66a9b0c84`
+- FlashInfer: 0.6.14
+
+The host's default `nvcc` was 11.8. The working CUDA toolkit came from the
+Python environment, so `CUDA_HOME` was set to
+`$VIRTUAL_ENV/lib/python3.12/site-packages/nvidia/cu13`.
+
+#### Command
+
+From the repository root, start the async-chunk server with:
+
+```bash
+export CUDA_HOME="$VIRTUAL_ENV/lib/python3.12/site-packages/nvidia/cu13"
+export PATH="$CUDA_HOME/bin:$PATH"
+export LD_LIBRARY_PATH="$CUDA_HOME/lib:${LD_LIBRARY_PATH:-}"
+
+CUDA_VISIBLE_DEVICES=0,1 vllm-omni serve zai-org/GLM-TTS \
+  --omni \
+  --trust-remote-code \
+  --port 8091 \
+  --deploy-config vllm_omni/deploy/glm_tts.yaml \
+  --stage-overrides '{"1":{"devices":"1"}}' \
+  --allowed-local-media-path "${PWD}/tests/assets"
+```
+
+Stage 0 runs on GPU 0, while the override places stage 1 on GPU 1. For the sync
+variant, add `--no-async-chunk` to the same command.
+
+#### Verification
+
+Run the Chinese voice-cloning check:
+
+```bash
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+  --model zai-org/GLM-TTS \
+  --text "今天天气很好，我们一起学习语音模型推理。" \
+  --ref-audio "file://${PWD}/tests/assets/glm_tts/jiayan_zh.wav" \
+  --ref-text "他当时还跟线下其他的站姐吵架，然后，打架进局子了。" \
+  --response-format wav \
+  --output glm_tts_zh.wav
+```
+
+Run the English voice-cloning check with the same Chinese reference:
+
+```bash
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+  --model zai-org/GLM-TTS \
+  --text "Hello, this is a voice cloning test with GLM-TTS." \
+  --ref-audio "file://${PWD}/tests/assets/glm_tts/jiayan_zh.wav" \
+  --ref-text "他当时还跟线下其他的站姐吵架，然后，打架进局子了。" \
+  --response-format wav \
+  --output glm_tts_en.wav
+```
+
+Inspect either output:
+
+```bash
+ffprobe -v error -select_streams a:0 \
+  -show_entries stream=codec_name,sample_rate,channels \
+  -of default=noprint_wrappers=1 glm_tts_zh.wav
+```
+
+Expected output:
+
+```text
+codec_name=pcm_s16le
+sample_rate=24000
+channels=1
+```
+
+The English smoke test produced a valid WAV with an actual duration of 5.020
+seconds.
+
+#### Observed warm-cache results
+
+These observations are from one controlled run, not a hardware or performance
+guarantee. Cold-start and JIT costs were excluded. Values are mean +/- standard
+deviation.
+
+##### Measurement procedure
+
+- For the single-GPU server, use the documented command with
+  `CUDA_VISIBLE_DEVICES=0` and omit `--stage-overrides`; both bundled stages
+  remain on GPU 0. For the dual-GPU server, use the command as shown.
+- Async chunking is the default. Add `--no-async-chunk` for sync mode.
+- For non-streaming measurements, use the exact Chinese client command above.
+  For streaming, add `--stream` to that command and change the output filename
+  to `glm_tts_zh.pcm`; streaming output is raw PCM, not WAV.
+- The fixed Chinese prompt and reference were exactly those shown in
+  Verification and yielded 3.900 seconds of audio. Each case used one warm-up
+  followed by three measured requests at concurrency 1.
+- Wall time starts immediately before the HTTP POST and stops after the complete
+  response body. TTFA runs from immediately before the POST to the first
+  non-empty audio bytes. RTF is wall time divided by audio duration.
+- Peak total device memory used was sampled every 0.2 seconds and reported by
+  NVML via `nvmlDeviceGetMemoryInfo.used`. The readings include background and
+  baseline use and are not baseline-subtracted.
+
+| GPUs | Mode | Response | Wall time (s) | TTFA (s) | RTF |
+|---:|---|---|---:|---:|---:|
+| 1 | Async | Non-streaming | 2.144 +/- 0.076 | -- | 0.550 +/- 0.019 |
+| 2 | Async | Non-streaming | 1.969 +/- 0.130 | -- | 0.505 +/- 0.033 |
+| 1 | Async | Streaming | 2.135 +/- 0.093 | 1.245 +/- 0.178 | 0.548 +/- 0.024 |
+| 2 | Async | Streaming | 1.891 +/- 0.006 | 1.170 +/- 0.042 | 0.485 +/- 0.002 |
+| 1 | Sync | Non-streaming | 1.822 +/- 0.016 | -- | 0.467 +/- 0.004 |
+| 2 | Sync | Non-streaming | 1.792 +/- 0.023 | -- | 0.459 +/- 0.006 |
+
+Moving stage 1 to GPU 1 reduced async non-streaming wall time by 8.2%. It
+reduced async streaming wall time by 11.4% and TTFA by 6.0%, and reduced sync
+wall time by 1.6%. Single- and dual-GPU outputs were bitwise identical only
+within matching execution and response modes.
+
+| Configuration | GPU 0 | GPU 1 |
+|---|---:|---:|
+| Single-GPU async | 31.52 GiB | -- |
+| Dual-GPU async | 29.23 GiB | 2.91 GiB |
+| Single-GPU sync | 31.47 GiB | -- |
+| Dual-GPU sync | 29.23 GiB | 2.91 GiB |