Commit
Β·
8d25b68
1
Parent(s):
6a9905a
secret
Browse files
README.md
CHANGED
@@ -18,12 +18,14 @@ Batch text classification using BERT-style encoder models (e.g., BERT, RoBERTa,
|
|
18 |
**Note**: This script is specifically for encoder-only classification models, not generative LLMs.
|
19 |
|
20 |
**Features:**
|
|
|
21 |
- π High-throughput batch processing
|
22 |
- π·οΈ Automatic label mapping from model config
|
23 |
- π Confidence scores for predictions
|
24 |
- π€ Direct integration with Hugging Face Hub
|
25 |
|
26 |
**Usage:**
|
|
|
27 |
```bash
|
28 |
# Local execution (requires GPU)
|
29 |
uv run classify-dataset.py \
|
@@ -35,6 +37,7 @@ uv run classify-dataset.py \
|
|
35 |
```
|
36 |
|
37 |
**HF Jobs execution:**
|
|
|
38 |
```bash
|
39 |
hf jobs uv run \
|
40 |
--flavor l4x1 \
|
@@ -52,6 +55,7 @@ hf jobs uv run \
|
|
52 |
Generate responses for chat-formatted prompts using generative LLMs (e.g., Llama, Qwen, Mistral) with vLLM's high-performance inference engine.
|
53 |
|
54 |
**Features:**
|
|
|
55 |
- π¬ Automatic chat template application
|
56 |
- π Multi-GPU tensor parallelism support
|
57 |
- π Smart filtering for prompts exceeding context length
|
@@ -60,6 +64,7 @@ Generate responses for chat-formatted prompts using generative LLMs (e.g., Llama
|
|
60 |
- ποΈ Full control over sampling parameters
|
61 |
|
62 |
**Usage:**
|
|
|
63 |
```bash
|
64 |
# Local execution with default Qwen model
|
65 |
uv run generate-responses.py \
|
@@ -79,12 +84,13 @@ uv run generate-responses.py \
|
|
79 |
```
|
80 |
|
81 |
**HF Jobs execution (multi-GPU):**
|
|
|
82 |
```bash
|
83 |
hf jobs uv run \
|
84 |
--flavor l4x4 \
|
85 |
--image vllm/vllm-openai \
|
86 |
-e UV_PRERELEASE=if-necessary \
|
87 |
-
-
|
88 |
https://huggingface.co/datasets/uv-scripts/vllm/raw/main/generate-responses.py \
|
89 |
davanstrien/cards_with_prompts \
|
90 |
davanstrien/test-generated-responses \
|
@@ -97,6 +103,7 @@ hf jobs uv run \
|
|
97 |
## π― Requirements
|
98 |
|
99 |
All scripts in this collection require:
|
|
|
100 |
- **NVIDIA GPU** with CUDA support
|
101 |
- **Python 3.10+**
|
102 |
- **UV package manager** ([install UV](https://docs.astral.sh/uv/getting-started/installation/))
|
@@ -104,6 +111,7 @@ All scripts in this collection require:
|
|
104 |
## π Performance Tips
|
105 |
|
106 |
### GPU Selection
|
|
|
107 |
- **L4 GPU** (`--flavor l4x1`): Best value for classification and smaller models
|
108 |
- **L4x4** (`--flavor l4x4`): Multi-GPU setup for large models (30B+ parameters)
|
109 |
- **A10 GPU** (`--flavor a10g-large`): Higher memory for larger models
|
@@ -111,16 +119,20 @@ All scripts in this collection require:
|
|
111 |
- Adjust batch size and tensor parallelism based on GPU configuration
|
112 |
|
113 |
### Batch Sizes
|
|
|
114 |
- **Classification**: Start with 10,000 locally, up to 100,000 on HF Jobs
|
115 |
- **Generation**: vLLM handles batching automatically - no manual configuration needed
|
116 |
|
117 |
### Multi-GPU Tensor Parallelism
|
|
|
118 |
- Auto-detects available GPUs by default
|
119 |
- Use `--tensor-parallel-size` to manually specify
|
120 |
- Required for models larger than single GPU memory (e.g., 30B+ models)
|
121 |
|
122 |
### Handling Long Contexts
|
|
|
123 |
The generate-responses.py script includes smart prompt filtering:
|
|
|
124 |
- **Default behavior**: Skips prompts exceeding max_model_len
|
125 |
- **Use `--max-model-len`**: Limit context to reduce memory usage
|
126 |
- **Use `--no-skip-long-prompts`**: Fail on long prompts instead of skipping
|
@@ -129,6 +141,7 @@ The generate-responses.py script includes smart prompt filtering:
|
|
129 |
## π About vLLM
|
130 |
|
131 |
vLLM is a high-throughput inference engine optimized for:
|
|
|
132 |
- Fast model serving with PagedAttention
|
133 |
- Efficient batch processing
|
134 |
- Support for various model architectures
|
@@ -137,13 +150,16 @@ vLLM is a high-throughput inference engine optimized for:
|
|
137 |
## π§ Technical Details
|
138 |
|
139 |
### UV Script Benefits
|
|
|
140 |
- **Zero setup**: Dependencies install automatically on first run
|
141 |
- **Reproducible**: Locked dependencies ensure consistent behavior
|
142 |
- **Self-contained**: Everything needed is in the script file
|
143 |
- **Direct execution**: Run from local files or URLs
|
144 |
|
145 |
### Dependencies
|
|
|
146 |
Scripts use UV's inline metadata for automatic dependency management:
|
|
|
147 |
```python
|
148 |
# /// script
|
149 |
# requires-python = ">=3.10"
|
@@ -161,15 +177,18 @@ Scripts use UV's inline metadata for automatic dependency management:
|
|
161 |
For bleeding-edge features, use the `UV_PRERELEASE=if-necessary` environment variable to allow pre-release versions when needed.
|
162 |
|
163 |
### Docker Image
|
|
|
164 |
For HF Jobs, we recommend the official vLLM Docker image: `vllm/vllm-openai`
|
165 |
|
166 |
This image includes:
|
|
|
167 |
- Pre-installed CUDA libraries
|
168 |
- vLLM and all dependencies
|
169 |
- UV package manager
|
170 |
- Optimized for GPU inference
|
171 |
|
172 |
### Environment Variables
|
|
|
173 |
- `HF_TOKEN`: Your Hugging Face authentication token (auto-detected if logged in)
|
174 |
- `UV_PRERELEASE=if-necessary`: Allow pre-release packages when required
|
175 |
- `HF_HUB_ENABLE_HF_TRANSFER=1`: Automatically enabled for faster downloads
|
@@ -177,6 +196,7 @@ This image includes:
|
|
177 |
## π Contributing
|
178 |
|
179 |
Have a vLLM script to share? We welcome contributions that:
|
|
|
180 |
- Solve real inference problems
|
181 |
- Include clear documentation
|
182 |
- Follow UV script best practices
|
@@ -186,4 +206,4 @@ Have a vLLM script to share? We welcome contributions that:
|
|
186 |
|
187 |
- [vLLM Documentation](https://docs.vllm.ai/)
|
188 |
- [UV Documentation](https://docs.astral.sh/uv/)
|
189 |
-
- [UV Scripts Organization](https://huggingface.co/uv-scripts)
|
|
|
18 |
**Note**: This script is specifically for encoder-only classification models, not generative LLMs.
|
19 |
|
20 |
**Features:**
|
21 |
+
|
22 |
- π High-throughput batch processing
|
23 |
- π·οΈ Automatic label mapping from model config
|
24 |
- π Confidence scores for predictions
|
25 |
- π€ Direct integration with Hugging Face Hub
|
26 |
|
27 |
**Usage:**
|
28 |
+
|
29 |
```bash
|
30 |
# Local execution (requires GPU)
|
31 |
uv run classify-dataset.py \
|
|
|
37 |
```
|
38 |
|
39 |
**HF Jobs execution:**
|
40 |
+
|
41 |
```bash
|
42 |
hf jobs uv run \
|
43 |
--flavor l4x1 \
|
|
|
55 |
Generate responses for chat-formatted prompts using generative LLMs (e.g., Llama, Qwen, Mistral) with vLLM's high-performance inference engine.
|
56 |
|
57 |
**Features:**
|
58 |
+
|
59 |
- π¬ Automatic chat template application
|
60 |
- π Multi-GPU tensor parallelism support
|
61 |
- π Smart filtering for prompts exceeding context length
|
|
|
64 |
- ποΈ Full control over sampling parameters
|
65 |
|
66 |
**Usage:**
|
67 |
+
|
68 |
```bash
|
69 |
# Local execution with default Qwen model
|
70 |
uv run generate-responses.py \
|
|
|
84 |
```
|
85 |
|
86 |
**HF Jobs execution (multi-GPU):**
|
87 |
+
|
88 |
```bash
|
89 |
hf jobs uv run \
|
90 |
--flavor l4x4 \
|
91 |
--image vllm/vllm-openai \
|
92 |
-e UV_PRERELEASE=if-necessary \
|
93 |
+
-s HF_TOKEN=hf_*** \
|
94 |
https://huggingface.co/datasets/uv-scripts/vllm/raw/main/generate-responses.py \
|
95 |
davanstrien/cards_with_prompts \
|
96 |
davanstrien/test-generated-responses \
|
|
|
103 |
## π― Requirements
|
104 |
|
105 |
All scripts in this collection require:
|
106 |
+
|
107 |
- **NVIDIA GPU** with CUDA support
|
108 |
- **Python 3.10+**
|
109 |
- **UV package manager** ([install UV](https://docs.astral.sh/uv/getting-started/installation/))
|
|
|
111 |
## π Performance Tips
|
112 |
|
113 |
### GPU Selection
|
114 |
+
|
115 |
- **L4 GPU** (`--flavor l4x1`): Best value for classification and smaller models
|
116 |
- **L4x4** (`--flavor l4x4`): Multi-GPU setup for large models (30B+ parameters)
|
117 |
- **A10 GPU** (`--flavor a10g-large`): Higher memory for larger models
|
|
|
119 |
- Adjust batch size and tensor parallelism based on GPU configuration
|
120 |
|
121 |
### Batch Sizes
|
122 |
+
|
123 |
- **Classification**: Start with 10,000 locally, up to 100,000 on HF Jobs
|
124 |
- **Generation**: vLLM handles batching automatically - no manual configuration needed
|
125 |
|
126 |
### Multi-GPU Tensor Parallelism
|
127 |
+
|
128 |
- Auto-detects available GPUs by default
|
129 |
- Use `--tensor-parallel-size` to manually specify
|
130 |
- Required for models larger than single GPU memory (e.g., 30B+ models)
|
131 |
|
132 |
### Handling Long Contexts
|
133 |
+
|
134 |
The generate-responses.py script includes smart prompt filtering:
|
135 |
+
|
136 |
- **Default behavior**: Skips prompts exceeding max_model_len
|
137 |
- **Use `--max-model-len`**: Limit context to reduce memory usage
|
138 |
- **Use `--no-skip-long-prompts`**: Fail on long prompts instead of skipping
|
|
|
141 |
## π About vLLM
|
142 |
|
143 |
vLLM is a high-throughput inference engine optimized for:
|
144 |
+
|
145 |
- Fast model serving with PagedAttention
|
146 |
- Efficient batch processing
|
147 |
- Support for various model architectures
|
|
|
150 |
## π§ Technical Details
|
151 |
|
152 |
### UV Script Benefits
|
153 |
+
|
154 |
- **Zero setup**: Dependencies install automatically on first run
|
155 |
- **Reproducible**: Locked dependencies ensure consistent behavior
|
156 |
- **Self-contained**: Everything needed is in the script file
|
157 |
- **Direct execution**: Run from local files or URLs
|
158 |
|
159 |
### Dependencies
|
160 |
+
|
161 |
Scripts use UV's inline metadata for automatic dependency management:
|
162 |
+
|
163 |
```python
|
164 |
# /// script
|
165 |
# requires-python = ">=3.10"
|
|
|
177 |
For bleeding-edge features, use the `UV_PRERELEASE=if-necessary` environment variable to allow pre-release versions when needed.
|
178 |
|
179 |
### Docker Image
|
180 |
+
|
181 |
For HF Jobs, we recommend the official vLLM Docker image: `vllm/vllm-openai`
|
182 |
|
183 |
This image includes:
|
184 |
+
|
185 |
- Pre-installed CUDA libraries
|
186 |
- vLLM and all dependencies
|
187 |
- UV package manager
|
188 |
- Optimized for GPU inference
|
189 |
|
190 |
### Environment Variables
|
191 |
+
|
192 |
- `HF_TOKEN`: Your Hugging Face authentication token (auto-detected if logged in)
|
193 |
- `UV_PRERELEASE=if-necessary`: Allow pre-release packages when required
|
194 |
- `HF_HUB_ENABLE_HF_TRANSFER=1`: Automatically enabled for faster downloads
|
|
|
196 |
## π Contributing
|
197 |
|
198 |
Have a vLLM script to share? We welcome contributions that:
|
199 |
+
|
200 |
- Solve real inference problems
|
201 |
- Include clear documentation
|
202 |
- Follow UV script best practices
|
|
|
206 |
|
207 |
- [vLLM Documentation](https://docs.vllm.ai/)
|
208 |
- [UV Documentation](https://docs.astral.sh/uv/)
|
209 |
+
- [UV Scripts Organization](https://huggingface.co/uv-scripts)
|