File size: 4,940 Bytes
a2ed5f0 1b0e1c0 a2ed5f0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 | ---
viewer: false
tags:
- uv-script
- transformers
- continuous-batching
- gpu
- inference
---
# Transformers Continuous Batching Scripts
GPU inference scripts using transformers' native continuous batching (CB). No vLLM dependency required.
## Why transformers CB?
- **Instant new model support** - works with any model supported by transformers, including newly released architectures. No waiting for vLLM to add support.
- **No dependency headaches** - no vLLM, flashinfer, or custom wheel indexes. Just `transformers` + `accelerate`.
- **Simple HF Jobs setup** - no Docker image needed. Just `hf jobs uv run`.
- **~95% of vLLM throughput** - uses PagedAttention and continuous scheduling for near-vLLM performance.
## Available Scripts
### generate-responses.py
Generate responses for prompts in a dataset. Supports chat messages and plain text prompts.
#### Quick Start
```bash
# Local (requires GPU)
uv run generate-responses.py \
username/input-dataset \
username/output-dataset \
--prompt-column question
# HF Jobs (single GPU)
hf jobs uv run --flavor l4x1 -s HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/transformers-inference/raw/main/generate-responses.py \
username/input-dataset \
username/output-dataset \
--prompt-column question \
--max-tokens 1024
# HF Jobs (multi-GPU for larger models)
hf jobs uv run --flavor l4x4 -s HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/transformers-inference/raw/main/generate-responses.py \
username/input-dataset \
username/output-dataset \
--model-id Qwen/Qwen3-30B-A3B-Instruct-2507 \
--messages-column messages \
--max-batch-tokens 2048 \
--max-tokens 4096
```
#### Example with SmolTalk2
```bash
# Generate responses for SmolTalk2 chat data
hf jobs uv run --flavor l4x1 -s HF_TOKEN \
https://huggingface.co/datasets/uv-scripts/transformers-inference/raw/main/generate-responses.py \
HuggingFaceTB/smoltalk2 username/smoltalk2-responses \
--subset SFT \
--split OpenHermes_2.5_no_think \
--messages-column messages \
--max-tokens 256
```
#### Parameters
| Parameter | Default | Description |
|-----------|---------|-------------|
| `--model-id` | `Qwen/Qwen3-4B-Instruct-2507` | Any HF causal LM model |
| `--messages-column` | `messages` | Column with chat messages |
| `--prompt-column` | - | Column with plain text prompts (alternative to messages) |
| `--output-column` | `response` | Name for the generated response column |
| `--temperature` | `0.7` | Sampling temperature |
| `--top-p` | `0.8` | Top-p (nucleus) sampling |
| `--top-k` | `20` | Top-k sampling |
| `--max-tokens` | `4096` | Maximum tokens to generate per response |
| `--repetition-penalty` | `1.0` | Repetition penalty |
| `--max-batch-tokens` | `512` | Token budget per scheduling step (see below) |
| `--dtype` | `bfloat16` | Model precision (`bfloat16`, `float16`, `float32`) |
| `--attn-implementation` | `paged|sdpa` | Attention backend (`paged|sdpa` or `paged\|flash_attention_2`) |
| `--max-samples` | all | Limit to N samples (useful for testing) |
| `--hf-token` | - | HF token (or use `HF_TOKEN` env var) |
| `--skip-long-prompts` | `True` | Skip prompts exceeding context length |
#### Tuning `--max-batch-tokens`
This is the key performance parameter. It controls how many tokens the continuous batching scheduler processes per step:
- **Too low** (e.g., 128): GPU underutilized, slow throughput
- **Too high** (e.g., 8192): May cause out-of-memory errors
- **Default 512**: Conservative, works on most GPUs
- **Recommended for A100/H100**: 2048-4096
- **Recommended for L4**: 512-1024
If you hit OOM errors, reduce this value or switch to `--dtype float16`.
## Current Limitations
- **Single GPU only** - `device_map="auto"` (pipeline parallelism) doesn't work with CB's PagedAttention cache. Transformers does have tensor parallelism (`tp_plan="auto"`) for supported models, but it requires `torchrun` and is undocumented with CB. For now, use a model that fits on one GPU (e.g., 8B in bf16 on A10G/L4 with 24GB).
- **Text-only** - no vision-language model support yet.
## When to use this vs vLLM
| | Transformers CB | vLLM |
|---|---|---|
| **Best for** | New/niche models, simple setup, avoiding dependency issues | Maximum throughput, production serving |
| **Model support** | Any transformers model, immediately | Popular models, may lag on new architectures |
| **Dependencies** | `transformers` + `accelerate` | `vllm` + `flashinfer` + custom indexes |
| **Docker image** | Not needed | `vllm/vllm-openai` recommended |
| **Multi-GPU** | Single GPU only (for now) | Tensor parallelism |
| **Performance** | ~95% of vLLM for text generation | Fastest for supported models |
| **VLM support** | Not yet | Yes |
**Rule of thumb**: Use transformers CB when you want simplicity and broad model support. Use vLLM when you need maximum throughput with well-supported models.
|