[ROCm][Recipe] Add MI300X coverage for verified recipes (#6207)
Signed-off-by: akshatvishu <akshatnayak197@gmail.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
akshatvishuandHongsheng Liu authored and GitHub committed
1 parent
309d9c3ea8
commit
e96d3fe2a6
7 files changed
+307
-6
No files matched your search
@@ -81,7 +81,7 @@ MammothModa2 generation parameters as a JSON object through `--extra-body`:
|
||||
```bash
|
||||
python examples/offline_inference/text_to_image/text_to_image.py \
|
||||
--model ./MammothModa2-Preview \
|
||||
--stage-configs-path vllm_omni/deploy/mammoth_moda2.yaml \
|
||||
--deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
|
||||
--prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
|
||||
--height 1024 \
|
||||
--width 1024 \
|
||||
@@ -132,6 +132,44 @@ ls -lh mammoth_t2i.png
|
||||
python -c "from PIL import Image; print(Image.open('mammoth_t2i.png').size)"
|
||||
```
|
||||
|
||||
### 1x AMD MI300X, MammothModa2 Preview
|
||||
|
||||
#### Environment
|
||||
|
||||
- OS: Linux 6.8.0-134-generic, x86_64
|
||||
- Container: official ROCm image built from `docker/Dockerfile.rocm`
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
|
||||
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
|
||||
- vLLM version: 0.27.0+rocm723
|
||||
- vLLM Omni version or commit: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
|
||||
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
|
||||
|
||||
#### Offline Commands
|
||||
|
||||
The checked run used the committed stage split, with `gpu_memory_utilization` set to 0.5 for AR and 0.3 for DiT:
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_image/text_to_image.py \
|
||||
--model bytedance-research/MammothModa2-Preview \
|
||||
--deploy-config vllm_omni/deploy/mammoth_moda2.yaml \
|
||||
--prompt "A stylish woman riding a motorcycle in NYC, movie poster style" \
|
||||
--height 1024 \
|
||||
--width 1024 \
|
||||
--seed 42 \
|
||||
--extra-body '{"text_guidance_scale": 4.0, "cfg_range": [0.0, 1.0], "num_inference_steps": 50}' \
|
||||
--enable-diffusion-pipeline-profiler \
|
||||
--log-stats \
|
||||
--output mammoth_t2i.png
|
||||
```
|
||||
|
||||
#### Verification
|
||||
|
||||
The first request took 85.224 seconds. The AR stage generated 4,161 visual tokens in 72.996 seconds, and the DiT stage took 12.163 seconds. AR weight loading used 21.4 GiB and took 8.250 seconds. DiT weight loading used 5.49 GiB and took 1.824 seconds. The largest one second whole device memory sample was 106.57 GiB, including the AR KV cache reserved by the 0.5 memory setting.
|
||||
|
||||
The output was a valid 1024 by 1024 RGB PNG.
|
||||
|
||||
## MammothModa2-Dev unified inference
|
||||
|
||||
MammothModa2-Dev uses a Qwen3-VL AR backbone, while MammothModa2-Preview uses
|
||||
|
||||
@@ -252,3 +252,53 @@ curl -X POST http://localhost:8091/v1/audio/speech \
|
||||
- stage_id: 1
|
||||
gpu_memory_utilization: 0.15
|
||||
```
|
||||
|
||||
### 1x AMD MI300X, 1.7B checkpoints
|
||||
|
||||
#### Environment
|
||||
|
||||
- OS: Linux 6.8.0-134-generic, x86_64
|
||||
- Container: official ROCm image built from `docker/Dockerfile.rocm`
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
|
||||
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
|
||||
- vLLM version: 0.27.0+rocm723
|
||||
- vLLM Omni version or commit: `3fecb6953ca8dc51210cc0421ef24552267a41ef`
|
||||
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
|
||||
- ONNX Runtime: onnxruntime-rocm 1.22.2.post3 with `ROCMExecutionProvider`
|
||||
- transformers: 5.15.0
|
||||
|
||||
#### Command
|
||||
|
||||
The MI300X tests also worked without `enforce_eager`. To use that setup, remove `enforce_eager: true` from stage 1 under `platforms.rocm` in `vllm_omni/deploy/qwen3_tts.yaml`.
|
||||
|
||||
Set this environment variable before you start Qwen3-TTS:
|
||||
|
||||
```bash
|
||||
export MIOPEN_FIND_MODE=FAST
|
||||
```
|
||||
|
||||
#### Verification
|
||||
|
||||
Each row records one offline request with one prompt and batch size one. Request time is the logged `e2e_wall_time_ms`, and stage 1 time is the logged `e2e_stage_1_wall_time_ms`.
|
||||
|
||||
| Checkpoint | Code2Wav configuration | `MIOPEN_FIND_MODE` | Request time | Stage 1 time | Output duration | Real time factor | Maximum sampled device memory |
|
||||
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: |
|
||||
| CustomVoice | Eager | Unset | 1.683 s | 1.613 s | 5.68 s | 0.30 | 60.91 GiB |
|
||||
| CustomVoice | Graphs allowed | `FAST` | 1.326 s | 1.259 s | 5.68 s | 0.23 | 62.19 GiB |
|
||||
| VoiceDesign | Eager | Unset | 1.195 s | 1.180 s | 4.40 s | 0.27 | 60.91 GiB |
|
||||
| VoiceDesign | Graphs allowed | `FAST` | 1.127 s | 1.054 s | 4.40 s | 0.26 | 62.21 GiB |
|
||||
| Base | Eager | Unset | 13.330 s | 13.263 s | 4.32 s | 3.09 | 61.74 GiB |
|
||||
| Base | Graphs allowed | `FAST` | 14.314 s | 14.245 s | 4.80 s | 2.98 | 62.84 GiB |
|
||||
|
||||
All six processes finished successfully, and each output passed the 24 kHz mono WAV checks.
|
||||
|
||||
#### Notes
|
||||
|
||||
- The current ROCm config keeps Code2Wav in eager mode.
|
||||
- The tested alternative allows graphs and sets [`MIOPEN_FIND_MODE=FAST`](https://rocm.docs.amd.com/projects/MIOpen/en/develop/reference/env_variables.html). MIOpen uses a saved kernel choice when available and uses its immediate fallback otherwise.
|
||||
- CustomVoice used a graph for one of four logged Code2Wav batches. VoiceDesign used a graph for two of four batches. Base used no graphs for its four logged batches.
|
||||
- The recorded request time with graphs allowed was lower for CustomVoice and VoiceDesign and higher for Base. The Base runs generated different amounts of audio, so their request times are not a direct comparison.
|
||||
- Each setup was tested with one request, so the timing is not a performance benchmark.
|
||||
- GPU memory was sampled once per second with `rocm-smi`.
|
||||
+4
-3
@@ -49,6 +49,7 @@ recipes/
|
||||
| [`Helios/Helios.md`](./Helios/Helios.md) | Text-to-video, image-to-video, and video-to-video generation | 1x NVIDIA H20 |
|
||||
| [`inclusionAI/Ming-flash-omni-2.0.md`](./inclusionAI/Ming-flash-omni-2.0.md) | Online serving for multimodal chat + standalone TTS | 4x H100 / 1x H100 80GB |
|
||||
| [`inclusionAI/Ming-omni-tts.md`](./inclusionAI/Ming-omni-tts.md) | Offline + online dense Ming TTS/audio generation | 1x H100 80GB / 1x AMD MI300X (ROCm 7.2) |
|
||||
| [`k2-fsa/OmniVoice.md`](./k2-fsa/OmniVoice.md) | Offline and online multilingual text-to-speech | 1x AMD MI300X (ROCm 7.2) |
|
||||
| [`IndexTeam/IndexTTS-2.md`](./IndexTeam/IndexTTS-2.md) | Online serving for voice-cloned TTS with optional emotion control | 1x L4 24GB or larger CUDA GPU |
|
||||
| [`IndexTeam/IndexTTS-2_5.md`](./IndexTeam/IndexTTS-2_5.md) | Online + offline multilingual voice-cloned TTS with native speed and emotion control | 1x NVIDIA H20 96GB |
|
||||
| [`MiniMaxAI/MiniMax-H3.md`](./MiniMaxAI/MiniMax-H3.md) | T2VA, FL2VA, image+audio Ref2VA, and multi-video Ref2VA serving | 1x GPU with CPU offload / 4x B300 |
|
||||
@@ -73,9 +74,9 @@ recipes/
|
||||
| [`Qwen/Qwen-Image-2512.md`](./Qwen/Qwen-Image-2512.md) | Text-to-image serving with step-wise continuous batching replay and ModelOpt FP8 / mixed FP8/NVFP4 | 1x A800 80GB / 2x B200 |
|
||||
| [`Qwen/Qwen-Image-Edit.md`](./Qwen/Qwen-Image-Edit.md) | Text-guided single-image editing | 1x or 2x H200 141GB |
|
||||
| [`Qwen/Qwen3-Omni.md`](./Qwen/Qwen3-Omni.md) | Online serving for multimodal chat | 1x A100 80GB |
|
||||
| [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB |
|
||||
| [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x H200 (144GB) |
|
||||
| [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB |
|
||||
| [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB / 1x AMD MI300X |
|
||||
| [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x or 2x H200 / 1x AMD MI300X |
|
||||
| [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB / 1x AMD MI300X |
|
||||
| [`Tencent/Covo-Audio-Chat.md`](./Tencent/Covo-Audio-Chat.md) | Online serving for audio chat | 1x A100 80GB |
|
||||
| [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |
|
||||
| [`Wan-AI/Wan2.2-T2V.md`](./Wan-AI/Wan2.2-T2V.md) | Text-to-video serving (Wan2.2 14B) with Skip-Softmax sparse attention | 1x datacenter Blackwell (B200/B300) |
|
||||
|
||||
@@ -178,6 +178,48 @@ with the expected 1536×2720 resolution.
|
||||
- The LLM transformer uses `QKVParallelLinear` and `MergedColumnParallelLinear`
|
||||
for fused QKV and gate/up projections with TP support.
|
||||
|
||||
### 1x AMD MI300X 192GB
|
||||
|
||||
#### Environment
|
||||
|
||||
- OS: Linux 6.8.0-134-generic, x86_64
|
||||
- Container: official ROCm image built from `docker/Dockerfile.rocm`
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
|
||||
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
|
||||
- vLLM version: 0.27.0+rocm723
|
||||
- vLLM Omni version or commit: `a704c8759c96e123c0d7c89b11f120b1c0f120cf`
|
||||
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
|
||||
|
||||
#### Command
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_image/text_to_image.py \
|
||||
--model SenseNova/SenseNova-U1-8B-MoT \
|
||||
--prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light, emotional documentary look." \
|
||||
--width 1536 \
|
||||
--height 2720 \
|
||||
--seed 42 \
|
||||
--num-inference-steps 50 \
|
||||
--cfg-scale 4.0 \
|
||||
--extra-body '{"think": true, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
|
||||
--enable-diffusion-pipeline-profiler \
|
||||
--log-stats \
|
||||
--output sensenova_u1_mi300x.png
|
||||
```
|
||||
|
||||
#### Verification
|
||||
|
||||
The command completed and wrote a valid 1536 by 2720 RGB PNG.
|
||||
|
||||
#### Notes
|
||||
|
||||
- Generation with 50 inference steps took 34.011 seconds.
|
||||
- Model loading used 32.774 GiB and took 13.987 seconds.
|
||||
- The internal profiler recorded 35.67 GB reserved and 35.10 GB allocated for the request.
|
||||
- The highest one second whole device memory sample was 37.93 GiB.
|
||||
|
||||
## Online Serving
|
||||
|
||||
SenseNova-U1 supports all four modalities via the OpenAI-compatible
|
||||
|
||||
@@ -164,7 +164,50 @@ PY
|
||||
`torchaudio==2.11.0+cu130`.
|
||||
- The `NIXL is not available`, `GLOO_SOCKET_IFNAME`, and `torchsde` boundary
|
||||
warnings observed during validation did not prevent successful generation.
|
||||
- This recipe was validated on one RTX 4090 24GB GPU only. Other GPU counts,
|
||||
ROCm, XPU, and NPU setups are not covered here.
|
||||
- The RTX 4090 entry was validated on one 24 GB GPU. The MI300X entry below covers one 192 GB GPU.
|
||||
- Long generations, higher inference-step counts, and non-WAV response formats
|
||||
were not benchmarked in this recipe.
|
||||
|
||||
#### 1x AMD MI300X 192GB
|
||||
|
||||
##### Environment
|
||||
|
||||
- OS: Linux 6.8.0-134-generic, x86_64
|
||||
- Container: official ROCm image built from `docker/Dockerfile.rocm`
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- Driver / runtime: AMD 6.19.14.31400000 / ROCm 7.2.53211
|
||||
- GPU: one AMD Instinct MI300X, `gfx942:sramecc+:xnack-`, 191.69 GiB visible HBM
|
||||
- vLLM version: 0.27.0+rocm723
|
||||
- vLLM Omni version or commit: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
|
||||
- Installed vLLM Omni package metadata: `0.27.0rc2.dev44+g55abdade9.rocm`
|
||||
|
||||
##### Commands
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_audio/text_to_audio.py \
|
||||
--model stabilityai/stable-audio-open-1.0 \
|
||||
--prompt "A gentle piano melody with soft room ambience" \
|
||||
--negative-prompt "Low quality, distorted, noisy" \
|
||||
--seed 42 \
|
||||
--guidance-scale 7.0 \
|
||||
--audio-length 10.0 \
|
||||
--num-inference-steps 50 \
|
||||
--cache-backend tea_cache \
|
||||
--enable-diffusion-pipeline-profiler \
|
||||
--output stable_audio_10s.wav
|
||||
```
|
||||
|
||||
##### Verification
|
||||
|
||||
The command completed and wrote a valid 44.1 kHz stereo WAV with 10.00 seconds of audio.
|
||||
|
||||
##### Notes
|
||||
|
||||
- TeaCache ran with `rel_l1_thresh=0.2`.
|
||||
- Model loading used 2.7891 GiB and took 3.706 seconds.
|
||||
- Generation took 4.750 seconds, which gives a real time factor of 0.475 for the 10.00 second output.
|
||||
- The internal profiler recorded 15.65 GB reserved and 9.68 GB allocated for the request.
|
||||
- The highest one second whole device memory sample was 19.61 GiB.
|
||||
- The output RMS was 0.0887, and its peak absolute amplitude was 0.5761.
|
||||
- The full process took 384 seconds, including startup and compilation.
|
||||
@@ -170,6 +170,35 @@ python examples/online_serving/text_to_speech/ming_tts/openai_speech_client.py \
|
||||
`--ref-audio` matches upstream `use_spk_emb=True`; do not add `--ref-text`
|
||||
for the dialect case.
|
||||
|
||||
**Offline smoke result on ROCm 7.2:**
|
||||
|
||||
A later offline smoke run used the same MI300X model with the following software:
|
||||
|
||||
- GPU: one AMD Instinct MI300X with 191.69 GiB visible HBM and gfx942
|
||||
- Kernel: Linux 6.8.0-134-generic, x86_64
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- ROCm or HIP: 7.2.53211
|
||||
- vLLM: 0.27.0+rocm723
|
||||
- vLLM Omni checkout: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_speech/ming_tts/end2end.py \
|
||||
--model inclusionAI/Ming-omni-tts-0.5B \
|
||||
--deploy-config vllm_omni/deploy/ming_tts.yaml \
|
||||
--case basic \
|
||||
--text "你好，这是 AMD MI300X ROCm 配方验证。" \
|
||||
--output-dir ./ming_tts_output \
|
||||
--output-name ming_omni_tts_mi300x.wav \
|
||||
--enforce-eager \
|
||||
--log-stats \
|
||||
--metadata-json ./ming_tts_manifest.json
|
||||
```
|
||||
|
||||
The checked run produced a valid 44.1 kHz mono WAV with 4.80 seconds of audio, RMS 0.1449, and peak absolute amplitude 0.8621. The first request took 20.406 seconds, which gives a real time factor of 4.25. The stage weight loads used 1.31 GiB and 1.47 GiB and took 1.756 seconds and 1.264 seconds. The largest one second whole device memory sample was 91.05 GiB, including the reserved KV cache. Both stages used eager mode, and the AR stage selected `TRITON_ATTN`.
|
||||
|
||||
The run used the official ROCm image built from `docker/Dockerfile.rocm`.
|
||||
|
||||
## Notes
|
||||
|
||||
- The official ROCm image includes the platform dependencies.
|
||||
|
||||
@@ -0,0 +1,98 @@
|
||||
# OmniVoice
|
||||
|
||||
> Offline and online text to speech with automatic voice, voice design, and reference voice cloning
|
||||
|
||||
## Summary
|
||||
|
||||
- Vendor: k2-fsa
|
||||
- Model: `k2-fsa/OmniVoice`
|
||||
- Task: Multilingual text to speech
|
||||
- Mode: Offline inference for all modes and online serving for automatic voice
|
||||
- Maintainer: Community
|
||||
|
||||
## References
|
||||
|
||||
- [Hugging Face model](https://huggingface.co/k2-fsa/OmniVoice)
|
||||
- [Offline text to speech examples](../../docs/user_guide/examples/offline_inference/text_to_speech.md#omnivoice)
|
||||
- [Online text to speech examples](../../docs/user_guide/examples/online_serving/text_to_speech.md#omnivoice)
|
||||
- [Deploy config](../../vllm_omni/deploy/omnivoice.yaml)
|
||||
|
||||
## Requirements
|
||||
|
||||
Download the checkpoint before the first run:
|
||||
|
||||
```bash
|
||||
huggingface-cli download k2-fsa/OmniVoice
|
||||
```
|
||||
|
||||
Automatic voice and voice design require `transformers>=4.57.0`. Reference voice cloning requires `transformers>=5.3.0`.
|
||||
|
||||
The vLLM Omni deploy config runs the OmniVoice generator and decoder inside one float32 diffusion stage.
|
||||
|
||||
## ROCm
|
||||
|
||||
### 1x AMD MI300X
|
||||
|
||||
#### Environment
|
||||
|
||||
- GPU: one AMD Instinct MI300X with 191.69 GiB visible HBM and gfx942
|
||||
- Kernel: Linux 6.8.0-134-generic, x86_64
|
||||
- Python: 3.12.13
|
||||
- PyTorch: 2.11.0+gitd0c8b1f
|
||||
- ROCm or HIP: 7.2.53211
|
||||
- vLLM: 0.27.0+rocm723
|
||||
- vLLM Omni checkout: `73e1368c7bb940efe1a025859c9d6c8eeeb2e3f0`
|
||||
- transformers: 5.15.0
|
||||
|
||||
The run used the official ROCm image built from `docker/Dockerfile.rocm`.
|
||||
|
||||
#### Automatic voice
|
||||
|
||||
Run the offline example from the repository root:
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
|
||||
--model k2-fsa/OmniVoice \
|
||||
--deploy-config vllm_omni/deploy/omnivoice.yaml \
|
||||
--text "Hello, this is OmniVoice running on one AMD MI300X." \
|
||||
--seed 42 \
|
||||
--output omnivoice_mi300x.wav
|
||||
```
|
||||
|
||||
The checked run produced a valid 24 kHz mono WAV with 3.60 seconds of audio, RMS 0.1194, and peak absolute amplitude 0.9129. The first request took 0.685 seconds, which gives a real time factor of 0.19. Model loading used 3.5739 GiB and took 9.564 seconds. The request used 4.56 GB reserved and 3.64 GB allocated according to the internal profiler, while the largest one second whole device sample was 8.33 GiB. The log states that the OmniVoice Triton kernels loaded.
|
||||
|
||||
#### Reference voice cloning
|
||||
|
||||
The offline interface also accepts a reference WAV and its transcript:
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
|
||||
--model k2-fsa/OmniVoice \
|
||||
--text "Hello, this is a cloned voice." \
|
||||
--ref-audio ref.wav \
|
||||
--ref-text "This is the reference transcription." \
|
||||
--output cloned.wav
|
||||
```
|
||||
|
||||
#### Voice design
|
||||
|
||||
```bash
|
||||
python3 examples/offline_inference/text_to_speech/omnivoice/end2end.py \
|
||||
--model k2-fsa/OmniVoice \
|
||||
--text "Hello, this is a designed voice." \
|
||||
--instruct "female, low pitch, British accent" \
|
||||
--output designed.wav
|
||||
```
|
||||
|
||||
## Online serving
|
||||
|
||||
Online serving currently supports automatic voice:
|
||||
|
||||
```bash
|
||||
vllm serve k2-fsa/OmniVoice --omni --port 8091 --trust-remote-code
|
||||
```
|
||||
|
||||
```bash
|
||||
cd examples/online_serving/text_to_speech/omnivoice
|
||||
python speech_client.py --text "Hello, how are you?"
|
||||
```
|
||||
Reference in new issue
Block a user