@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
md 116 lines 4.85 KB
Rendered Raw
1 # Offline Inference
2
3 Offline inference is possible in your own code using vLLM's [`LLM`][vllm.LLM] class.
4
5 ## Model Types
6
7 vLLM models can be categorized into two types:
8
9 - **[Generative Models](../models/supported_models.md)** - Models that produce text completions or chat responses (e.g., LLaMA, Qwen, DeepSeek). Use `LLM.generate()` and `LLM.chat()` for these models.
10
11 - **[Pooling Models](../models/pooling_models/README.md)** - These models do not generate content. They are primarily used for classification and retrieval tasks, such as bge-m3 and Qwen3 Reranker.
12
13 ## Generative APIs
14
15 For further details on generative models, please refer to [this page](../models/supported_models.md).
16
17 - `LLM.generate` - Generates completions for the given input prompts.
18 - `LLM.chat` - Generates responses for a chat conversation.
19
20 ## Asynchronous Queue APIs
21
22 - `LLM.enqueue` - Enqueues prompts for generation without waiting for completion.
23 - `LLM.enqueue_chat` - Enqueues chat conversations for generation without waiting.
24 - `LLM.wait_for_completion` - Waits for all enqueued requests to complete and returns results.
25
26 ## Pooling APIs
27
28 For further details on pooling models, please refer to [this page](../models/pooling_models/README.md).
29
30 - `LLM.classify` - Only applicable to [classification models](../models/pooling_models/classify.md).
31 - `LLM.embed` - Only applicable to [embedding models](../models/pooling_models/embed.md).
32 - `LLM.score` - Applicable to [score models](../models/pooling_models/scoring.md) (cross-encoder, bi-encoder, late-interaction).
33 - `LLM.encode` - Applicable to all [pooling models](../models/pooling_models/README.md).
34
35 ## Profiling APIs
36
37 For further details on profiling, please refer to [this page](../contributing/profiling.md).
38
39 - `LLM.start_profile` - Starts profiling with an optional custom trace prefix.
40 - `LLM.stop_profile` - Stops the ongoing profiling session.
41
42 ## Sleep Mode APIs
43
44 For further details on sleep mode, please refer to [this page](../features/sleep_mode.md).
45
46 - `LLM.sleep` - Puts the engine into sleep mode.
47 - `LLM.wake_up` - Wakes up the engine from sleep mode.
48
49 ## Cache Management APIs
50
51 - `LLM.reset_mm_cache` - Resets the multi-modal cache.
52 - `LLM.reset_prefix_cache` - Resets the prefix cache.
53
54 ## Metrics APIs
55
56 For further details on metrics, please refer to [this page](../design/metrics.md).
57
58 - `LLM.get_metrics` - Returns a snapshot of aggregated metrics from Prometheus.
59
60 ## Weight Transfer APIs (RL Training)
61
62 For further details on Weight Transfer, please refer to [this page](../training/weight_transfer/README.md).
63
64 - `LLM.init_weight_transfer_engine` - Initializes the weight transfer engine for RL training.
65 - `LLM.start_weight_update` - Starts a new weight update cycle.
66 - `LLM.update_weights` - Updates the model weights.
67 - `LLM.finish_weight_update` - Finishes the current weight update cycle.
68 - `LLM.update_weight_version` - Sets the weight version without updating model weights.
69 - `LLM.get_weight_version` - Returns the latest committed weight version.
70
71 ## Additional APIs
72
73 - `LLM.collective_rpc` - Executes a method or callable collectively across all workers.
74 - `LLM.apply_model` - Applies a function directly to the model inside each worker.
75
76 ## API Reference
77
78 [Offline Inference](../api/README.md#offline-inference)
79
80 ## Ray Data LLM API
81
82 Ray Data LLM is an alternative offline inference API that uses vLLM as the underlying engine.
83 This API adds several batteries-included capabilities that simplify large-scale, GPU-efficient inference:
84
85 - Streaming execution processes datasets that exceed aggregate cluster memory.
86 - Automatic sharding, load balancing, and autoscaling distribute work across a Ray cluster with built-in fault tolerance.
87 - Continuous batching keeps vLLM replicas saturated and maximizes GPU utilization.
88 - Transparent support for tensor and pipeline parallelism enables efficient multi-GPU inference.
89 - Reading and writing to most popular file formats and cloud object storage.
90 - Scaling up the workload without code changes.
91
92 ??? code
93
94 ```python
95 import ray # Requires ray>=2.44.1
96 from ray.data.llm import vLLMEngineProcessorConfig, build_llm_processor
97
98 config = vLLMEngineProcessorConfig(model_source="unsloth/Llama-3.2-1B-Instruct")
99 processor = build_llm_processor(
100 config,
101 preprocess=lambda row: {
102 "messages": [
103 {"role": "system", "content": "You are a bot that completes unfinished haikus."},
104 {"role": "user", "content": row["item"]},
105 ],
106 "sampling_params": {"temperature": 0.3, "max_tokens": 250},
107 },
108 postprocess=lambda row: {"answer": row["generated_text"]},
109 )
110
111 ds = ray.data.from_items(["An old silent pond..."])
112 ds = processor(ds)
113 ds.write_parquet("local:///tmp/data/")
114 ```
115
116 For more information about the Ray Data LLM API, see the [Ray Data LLM documentation](https://docs.ray.io/en/latest/data/working-with-llms.html).