[Model]: Add transformers backend support

[ArthurZucker]

Adds support for transformers as a backend

Following huggingface/transformers#35235, a bunch of models should already be supported, we are ramping up support for more models.

Thanks @Isotr0py for the TP support, and @hmellor for his help as well!
This includes:

• trust_remote_code=True support: any model on the hub, if it implements attention the correct way can be natively supported!!
• tensor parallel support

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[ywang96]

Hello @ArthurZucker! This is very exciting!

I know this PR is still a draft, but could you provide some context on the scope of this effort? Is it to support any model on transformers?

[ArthurZucker]

Yep, overall this should support any model that is supported in transformers, were the cache is "simple" so for now, most of the decoder models and the encoder models for a single modularity!
For MultiModal models, we might need a little bit of extra work, but I thing LLAVA models should work out of the box!

We are refactor our models to make sure it's propagated to as many models as possible!

[ArthurZucker]

Might not have time to finish this week, will make it ready for next week 🎄
This should be minimal (no support fort Lora or at least I am not testing it ! This might. need to either call transformers's from pretrain, or replace modules similarly to TP)

[hmellor]

Benchmarks on A100 using the following command:

python benchmarks/benchmark_throughput.py --backend vllm --model meta-llama/Llama-3.1-8B-Instruct --dataset ShareGPT_V3_unfiltered_cleaned_split.json

Results:

Class	Result
LlamaForCausalLM	Throughput: 12.88 requests/s, 5325.05 total tokens/s, 2554.02 output tokens/s
TransformersModel	Throughput: 11.38 requests/s, 4705.90 total tokens/s, 2257.06 output tokens/s

[Isotr0py]

transformers backend should work with V1 now:

INFO 02-01 14:05:23 __init__.py:183] Automatically detected platform cuda.
WARNING 02-01 14:05:25 arg_utils.py:1325] Setting max_num_batched_tokens to 8192 for LLM_CLASS usage context.
INFO 02-01 14:05:31 config.py:540] This model supports multiple tasks: {'classify', 'score', 'reward', 'embed', 'generate'}. Defaulting to 'generate'.
INFO 02-01 14:05:31 config.py:1508] Chunked prefill is enabled with max_num_batched_tokens=8192.
INFO 02-01 14:05:32 core.py:45] Initializing a V1 LLM engine (v0.1.dev4338+g2079e43) with config: model='../Llama-3.2-1B-Instruct', speculative_config=None, tokenizer='../Llama-3.2-1B-Instruct', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=131072, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='xgrammar'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=../Llama-3.2-1B-Instruct, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, disable_mm_preprocessor_cache=False, mm_processor_kwargs=None, pooler_config=None, compilation_config={"level":3,"custom_ops":["none"],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output"],"use_inductor":true,"compile_sizes":[],"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":512}
WARNING 02-01 14:05:33 registry.py:336] `mm_limits` has already been set for model=../Llama-3.2-1B-Instruct, and will be overwritten by the new values.
INFO 02-01 14:05:33 gpu_model_runner.py:843] Starting to load model ../Llama-3.2-1B-Instruct...
INFO 02-01 14:05:33 transformers.py:129] Using Transformers backend.
INFO 02-01 14:05:33 cuda.py:162] Using Flash Attention backend on V1 engine.
WARNING 02-01 14:05:33 topk_topp_sampler.py:44] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
WARNING 02-01 14:05:33 config.py:3361] `torch.compile` is turned on, but the model ../Llama-3.2-1B-Instruct does not support it. Please open an issue on GitHubif you want it to be supported.
Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:02<00:00,  2.40s/it]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:02<00:00,  2.40s/it]

INFO 02-01 14:05:36 gpu_model_runner.py:848] Loading model weights took 2.3029 GB
INFO 02-01 14:05:37 kv_cache_utils.py:395] # GPU blocks: 11864
INFO 02-01 14:05:39 gpu_model_runner.py:1019] Graph capturing finished in 2 secs, took 0.09 GiB
INFO 02-01 14:05:39 core.py:89] init engine (profile, create kv cache, warmup model) took 3.11 seconds
Processed prompts: 100%|████████████████████████████████████████████████| 4/4 [00:00<00:00,  7.99it/s, est. speed input: 51.99 toks/s, output: 127.97 toks/s]
Prompt: 'Hello, my name is', Generated text: ' Emma and I am a 4th grade student. I just wanted to say'
Prompt: 'The president of the United States is', Generated text: ' the head of state and the head of government. He or she is typically the'
Prompt: 'The capital of France is', Generated text: ' Paris. The French word for the phrase "good morning" is "bonjour'
Prompt: 'The future of AI is', Generated text: ' multifaceted, and researchers are exploring various applications that will benefit society. One'

[Isotr0py]

Let's get this merged first! We can add BNB and LoRA support in other following PR.

[DarkLight1337]

Please fix the failing tests

[DarkLight1337]

Please also add the distributed transformers test to the distributed tests CI